1580158
/
sd-webui-controlnet
mirror of https://github.com/Mikubill/sd-webui-controlnet.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

sdxl support (for SAI/HuggingFace/diffuser/community models) (#1952) 

sdxl support (for SAI/HuggingFace/diffuser/community models)
1.1.400
pull/2041/head
lllyasviel 2023-09-04 10:41:42 -07:00 committed by GitHub
parent 664ac74cae
commit 1d5402326d
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
68 changed files with 2649 additions and 150192 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

687
LICENSE
View File

@ -1,21 +1,674 @@
MIT License
GNU GENERAL PUBLIC LICENSE
Version 3, 29 June 2007
Copyright (c) 2023 Kakigōri Maker
Copyright (C) 2007 Free Software Foundation, Inc. <https://fsf.org/>
Everyone is permitted to copy and distribute verbatim copies
of this license document, but changing it is not allowed.
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
Preamble
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
The GNU General Public License is a free, copyleft license for
software and other kinds of works.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
The licenses for most software and other practical works are designed
to take away your freedom to share and change the works. By contrast,
the GNU General Public License is intended to guarantee your freedom to
share and change all versions of a program--to make sure it remains free
software for all its users. We, the Free Software Foundation, use the
GNU General Public License for most of our software; it applies also to
any other work released this way by its authors. You can apply it to
your programs, too.
When we speak of free software, we are referring to freedom, not
price. Our General Public Licenses are designed to make sure that you
have the freedom to distribute copies of free software (and charge for
them if you wish), that you receive source code or can get it if you
want it, that you can change the software or use pieces of it in new
free programs, and that you know you can do these things.
To protect your rights, we need to prevent others from denying you
these rights or asking you to surrender the rights. Therefore, you have
certain responsibilities if you distribute copies of the software, or if
you modify it: responsibilities to respect the freedom of others.
For example, if you distribute copies of such a program, whether
gratis or for a fee, you must pass on to the recipients the same
freedoms that you received. You must make sure that they, too, receive
or can get the source code. And you must show them these terms so they
know their rights.
Developers that use the GNU GPL protect your rights with two steps:
(1) assert copyright on the software, and (2) offer you this License
giving you legal permission to copy, distribute and/or modify it.
For the developers' and authors' protection, the GPL clearly explains
that there is no warranty for this free software. For both users' and
authors' sake, the GPL requires that modified versions be marked as
changed, so that their problems will not be attributed erroneously to
authors of previous versions.
Some devices are designed to deny users access to install or run
modified versions of the software inside them, although the manufacturer
can do so. This is fundamentally incompatible with the aim of
protecting users' freedom to change the software. The systematic
pattern of such abuse occurs in the area of products for individuals to
use, which is precisely where it is most unacceptable. Therefore, we
have designed this version of the GPL to prohibit the practice for those
products. If such problems arise substantially in other domains, we
stand ready to extend this provision to those domains in future versions
of the GPL, as needed to protect the freedom of users.
Finally, every program is threatened constantly by software patents.
States should not allow patents to restrict development and use of
software on general-purpose computers, but in those that do, we wish to
avoid the special danger that patents applied to a free program could
make it effectively proprietary. To prevent this, the GPL assures that
patents cannot be used to render the program non-free.
The precise terms and conditions for copying, distribution and
modification follow.
TERMS AND CONDITIONS
0. Definitions.
"This License" refers to version 3 of the GNU General Public License.
"Copyright" also means copyright-like laws that apply to other kinds of
works, such as semiconductor masks.
"The Program" refers to any copyrightable work licensed under this
License. Each licensee is addressed as "you". "Licensees" and
"recipients" may be individuals or organizations.
To "modify" a work means to copy from or adapt all or part of the work
in a fashion requiring copyright permission, other than the making of an
exact copy. The resulting work is called a "modified version" of the
earlier work or a work "based on" the earlier work.
A "covered work" means either the unmodified Program or a work based
on the Program.
To "propagate" a work means to do anything with it that, without
permission, would make you directly or secondarily liable for
infringement under applicable copyright law, except executing it on a
computer or modifying a private copy. Propagation includes copying,
distribution (with or without modification), making available to the
public, and in some countries other activities as well.
To "convey" a work means any kind of propagation that enables other
parties to make or receive copies. Mere interaction with a user through
a computer network, with no transfer of a copy, is not conveying.
An interactive user interface displays "Appropriate Legal Notices"
to the extent that it includes a convenient and prominently visible
feature that (1) displays an appropriate copyright notice, and (2)
tells the user that there is no warranty for the work (except to the
extent that warranties are provided), that licensees may convey the
work under this License, and how to view a copy of this License. If
the interface presents a list of user commands or options, such as a
menu, a prominent item in the list meets this criterion.
1. Source Code.
The "source code" for a work means the preferred form of the work
for making modifications to it. "Object code" means any non-source
form of a work.
A "Standard Interface" means an interface that either is an official
standard defined by a recognized standards body, or, in the case of
interfaces specified for a particular programming language, one that
is widely used among developers working in that language.
The "System Libraries" of an executable work include anything, other
than the work as a whole, that (a) is included in the normal form of
packaging a Major Component, but which is not part of that Major
Component, and (b) serves only to enable use of the work with that
Major Component, or to implement a Standard Interface for which an
implementation is available to the public in source code form. A
"Major Component", in this context, means a major essential component
(kernel, window system, and so on) of the specific operating system
(if any) on which the executable work runs, or a compiler used to
produce the work, or an object code interpreter used to run it.
The "Corresponding Source" for a work in object code form means all
the source code needed to generate, install, and (for an executable
work) run the object code and to modify the work, including scripts to
control those activities. However, it does not include the work's
System Libraries, or general-purpose tools or generally available free
programs which are used unmodified in performing those activities but
which are not part of the work. For example, Corresponding Source
includes interface definition files associated with source files for
the work, and the source code for shared libraries and dynamically
linked subprograms that the work is specifically designed to require,
such as by intimate data communication or control flow between those
subprograms and other parts of the work.
The Corresponding Source need not include anything that users
can regenerate automatically from other parts of the Corresponding
Source.
The Corresponding Source for a work in source code form is that
same work.
2. Basic Permissions.
All rights granted under this License are granted for the term of
copyright on the Program, and are irrevocable provided the stated
conditions are met. This License explicitly affirms your unlimited
permission to run the unmodified Program. The output from running a
covered work is covered by this License only if the output, given its
content, constitutes a covered work. This License acknowledges your
rights of fair use or other equivalent, as provided by copyright law.
You may make, run and propagate covered works that you do not
convey, without conditions so long as your license otherwise remains
in force. You may convey covered works to others for the sole purpose
of having them make modifications exclusively for you, or provide you
with facilities for running those works, provided that you comply with
the terms of this License in conveying all material for which you do
not control copyright. Those thus making or running the covered works
for you must do so exclusively on your behalf, under your direction
and control, on terms that prohibit them from making any copies of
your copyrighted material outside their relationship with you.
Conveying under any other circumstances is permitted solely under
the conditions stated below. Sublicensing is not allowed; section 10
makes it unnecessary.
3. Protecting Users' Legal Rights From Anti-Circumvention Law.
No covered work shall be deemed part of an effective technological
measure under any applicable law fulfilling obligations under article
11 of the WIPO copyright treaty adopted on 20 December 1996, or
similar laws prohibiting or restricting circumvention of such
measures.
When you convey a covered work, you waive any legal power to forbid
circumvention of technological measures to the extent such circumvention
is effected by exercising rights under this License with respect to
the covered work, and you disclaim any intention to limit operation or
modification of the work as a means of enforcing, against the work's
users, your or third parties' legal rights to forbid circumvention of
technological measures.
4. Conveying Verbatim Copies.
You may convey verbatim copies of the Program's source code as you
receive it, in any medium, provided that you conspicuously and
appropriately publish on each copy an appropriate copyright notice;
keep intact all notices stating that this License and any
non-permissive terms added in accord with section 7 apply to the code;
keep intact all notices of the absence of any warranty; and give all
recipients a copy of this License along with the Program.
You may charge any price or no price for each copy that you convey,
and you may offer support or warranty protection for a fee.
5. Conveying Modified Source Versions.
You may convey a work based on the Program, or the modifications to
produce it from the Program, in the form of source code under the
terms of section 4, provided that you also meet all of these conditions:
a) The work must carry prominent notices stating that you modified
it, and giving a relevant date.
b) The work must carry prominent notices stating that it is
released under this License and any conditions added under section
7. This requirement modifies the requirement in section 4 to
"keep intact all notices".
c) You must license the entire work, as a whole, under this
License to anyone who comes into possession of a copy. This
License will therefore apply, along with any applicable section 7
additional terms, to the whole of the work, and all its parts,
regardless of how they are packaged. This License gives no
permission to license the work in any other way, but it does not
invalidate such permission if you have separately received it.
d) If the work has interactive user interfaces, each must display
Appropriate Legal Notices; however, if the Program has interactive
interfaces that do not display Appropriate Legal Notices, your
work need not make them do so.
A compilation of a covered work with other separate and independent
works, which are not by their nature extensions of the covered work,
and which are not combined with it such as to form a larger program,
in or on a volume of a storage or distribution medium, is called an
"aggregate" if the compilation and its resulting copyright are not
used to limit the access or legal rights of the compilation's users
beyond what the individual works permit. Inclusion of a covered work
in an aggregate does not cause this License to apply to the other
parts of the aggregate.
6. Conveying Non-Source Forms.
You may convey a covered work in object code form under the terms
of sections 4 and 5, provided that you also convey the
machine-readable Corresponding Source under the terms of this License,
in one of these ways:
a) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by the
Corresponding Source fixed on a durable physical medium
customarily used for software interchange.
b) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by a
written offer, valid for at least three years and valid for as
long as you offer spare parts or customer support for that product
model, to give anyone who possesses the object code either (1) a
copy of the Corresponding Source for all the software in the
product that is covered by this License, on a durable physical
medium customarily used for software interchange, for a price no
more than your reasonable cost of physically performing this
conveying of source, or (2) access to copy the
Corresponding Source from a network server at no charge.
c) Convey individual copies of the object code with a copy of the
written offer to provide the Corresponding Source. This
alternative is allowed only occasionally and noncommercially, and
only if you received the object code with such an offer, in accord
with subsection 6b.
d) Convey the object code by offering access from a designated
place (gratis or for a charge), and offer equivalent access to the
Corresponding Source in the same way through the same place at no
further charge. You need not require recipients to copy the
Corresponding Source along with the object code. If the place to
copy the object code is a network server, the Corresponding Source
may be on a different server (operated by you or a third party)
that supports equivalent copying facilities, provided you maintain
clear directions next to the object code saying where to find the
Corresponding Source. Regardless of what server hosts the
Corresponding Source, you remain obligated to ensure that it is
available for as long as needed to satisfy these requirements.
e) Convey the object code using peer-to-peer transmission, provided
you inform other peers where the object code and Corresponding
Source of the work are being offered to the general public at no
charge under subsection 6d.
A separable portion of the object code, whose source code is excluded
from the Corresponding Source as a System Library, need not be
included in conveying the object code work.
A "User Product" is either (1) a "consumer product", which means any
tangible personal property which is normally used for personal, family,
or household purposes, or (2) anything designed or sold for incorporation
into a dwelling. In determining whether a product is a consumer product,
doubtful cases shall be resolved in favor of coverage. For a particular
product received by a particular user, "normally used" refers to a
typical or common use of that class of product, regardless of the status
of the particular user or of the way in which the particular user
actually uses, or expects or is expected to use, the product. A product
is a consumer product regardless of whether the product has substantial
commercial, industrial or non-consumer uses, unless such uses represent
the only significant mode of use of the product.
"Installation Information" for a User Product means any methods,
procedures, authorization keys, or other information required to install
and execute modified versions of a covered work in that User Product from
a modified version of its Corresponding Source. The information must
suffice to ensure that the continued functioning of the modified object
code is in no case prevented or interfered with solely because
modification has been made.
If you convey an object code work under this section in, or with, or
specifically for use in, a User Product, and the conveying occurs as
part of a transaction in which the right of possession and use of the
User Product is transferred to the recipient in perpetuity or for a
fixed term (regardless of how the transaction is characterized), the
Corresponding Source conveyed under this section must be accompanied
by the Installation Information. But this requirement does not apply
if neither you nor any third party retains the ability to install
modified object code on the User Product (for example, the work has
been installed in ROM).
The requirement to provide Installation Information does not include a
requirement to continue to provide support service, warranty, or updates
for a work that has been modified or installed by the recipient, or for
the User Product in which it has been modified or installed. Access to a
network may be denied when the modification itself materially and
adversely affects the operation of the network or violates the rules and
protocols for communication across the network.
Corresponding Source conveyed, and Installation Information provided,
in accord with this section must be in a format that is publicly
documented (and with an implementation available to the public in
source code form), and must require no special password or key for
unpacking, reading or copying.
7. Additional Terms.
"Additional permissions" are terms that supplement the terms of this
License by making exceptions from one or more of its conditions.
Additional permissions that are applicable to the entire Program shall
be treated as though they were included in this License, to the extent
that they are valid under applicable law. If additional permissions
apply only to part of the Program, that part may be used separately
under those permissions, but the entire Program remains governed by
this License without regard to the additional permissions.
When you convey a copy of a covered work, you may at your option
remove any additional permissions from that copy, or from any part of
it. (Additional permissions may be written to require their own
removal in certain cases when you modify the work.) You may place
additional permissions on material, added by you to a covered work,
for which you have or can give appropriate copyright permission.
Notwithstanding any other provision of this License, for material you
add to a covered work, you may (if authorized by the copyright holders of
that material) supplement the terms of this License with terms:
a) Disclaiming warranty or limiting liability differently from the
terms of sections 15 and 16 of this License; or
b) Requiring preservation of specified reasonable legal notices or
author attributions in that material or in the Appropriate Legal
Notices displayed by works containing it; or
c) Prohibiting misrepresentation of the origin of that material, or
requiring that modified versions of such material be marked in
reasonable ways as different from the original version; or
d) Limiting the use for publicity purposes of names of licensors or
authors of the material; or
e) Declining to grant rights under trademark law for use of some
trade names, trademarks, or service marks; or
f) Requiring indemnification of licensors and authors of that
material by anyone who conveys the material (or modified versions of
it) with contractual assumptions of liability to the recipient, for
any liability that these contractual assumptions directly impose on
those licensors and authors.
All other non-permissive additional terms are considered "further
restrictions" within the meaning of section 10. If the Program as you
received it, or any part of it, contains a notice stating that it is
governed by this License along with a term that is a further
restriction, you may remove that term. If a license document contains
a further restriction but permits relicensing or conveying under this
License, you may add to a covered work material governed by the terms
of that license document, provided that the further restriction does
not survive such relicensing or conveying.
If you add terms to a covered work in accord with this section, you
must place, in the relevant source files, a statement of the
additional terms that apply to those files, or a notice indicating
where to find the applicable terms.
Additional terms, permissive or non-permissive, may be stated in the
form of a separately written license, or stated as exceptions;
the above requirements apply either way.
8. Termination.
You may not propagate or modify a covered work except as expressly
provided under this License. Any attempt otherwise to propagate or
modify it is void, and will automatically terminate your rights under
this License (including any patent licenses granted under the third
paragraph of section 11).
However, if you cease all violation of this License, then your
license from a particular copyright holder is reinstated (a)
provisionally, unless and until the copyright holder explicitly and
finally terminates your license, and (b) permanently, if the copyright
holder fails to notify you of the violation by some reasonable means
prior to 60 days after the cessation.
Moreover, your license from a particular copyright holder is
reinstated permanently if the copyright holder notifies you of the
violation by some reasonable means, this is the first time you have
received notice of violation of this License (for any work) from that
copyright holder, and you cure the violation prior to 30 days after
your receipt of the notice.
Termination of your rights under this section does not terminate the
licenses of parties who have received copies or rights from you under
this License. If your rights have been terminated and not permanently
reinstated, you do not qualify to receive new licenses for the same
material under section 10.
9. Acceptance Not Required for Having Copies.
You are not required to accept this License in order to receive or
run a copy of the Program. Ancillary propagation of a covered work
occurring solely as a consequence of using peer-to-peer transmission
to receive a copy likewise does not require acceptance. However,
nothing other than this License grants you permission to propagate or
modify any covered work. These actions infringe copyright if you do
not accept this License. Therefore, by modifying or propagating a
covered work, you indicate your acceptance of this License to do so.
10. Automatic Licensing of Downstream Recipients.
Each time you convey a covered work, the recipient automatically
receives a license from the original licensors, to run, modify and
propagate that work, subject to this License. You are not responsible
for enforcing compliance by third parties with this License.
An "entity transaction" is a transaction transferring control of an
organization, or substantially all assets of one, or subdividing an
organization, or merging organizations. If propagation of a covered
work results from an entity transaction, each party to that
transaction who receives a copy of the work also receives whatever
licenses to the work the party's predecessor in interest had or could
give under the previous paragraph, plus a right to possession of the
Corresponding Source of the work from the predecessor in interest, if
the predecessor has it or can get it with reasonable efforts.
You may not impose any further restrictions on the exercise of the
rights granted or affirmed under this License. For example, you may
not impose a license fee, royalty, or other charge for exercise of
rights granted under this License, and you may not initiate litigation
(including a cross-claim or counterclaim in a lawsuit) alleging that
any patent claim is infringed by making, using, selling, offering for
sale, or importing the Program or any portion of it.
11. Patents.
A "contributor" is a copyright holder who authorizes use under this
License of the Program or a work on which the Program is based. The
work thus licensed is called the contributor's "contributor version".
A contributor's "essential patent claims" are all patent claims
owned or controlled by the contributor, whether already acquired or
hereafter acquired, that would be infringed by some manner, permitted
by this License, of making, using, or selling its contributor version,
but do not include claims that would be infringed only as a
consequence of further modification of the contributor version. For
purposes of this definition, "control" includes the right to grant
patent sublicenses in a manner consistent with the requirements of
this License.
Each contributor grants you a non-exclusive, worldwide, royalty-free
patent license under the contributor's essential patent claims, to
make, use, sell, offer for sale, import and otherwise run, modify and
propagate the contents of its contributor version.
In the following three paragraphs, a "patent license" is any express
agreement or commitment, however denominated, not to enforce a patent
(such as an express permission to practice a patent or covenant not to
sue for patent infringement). To "grant" such a patent license to a
party means to make such an agreement or commitment not to enforce a
patent against the party.
If you convey a covered work, knowingly relying on a patent license,
and the Corresponding Source of the work is not available for anyone
to copy, free of charge and under the terms of this License, through a
publicly available network server or other readily accessible means,
then you must either (1) cause the Corresponding Source to be so
available, or (2) arrange to deprive yourself of the benefit of the
patent license for this particular work, or (3) arrange, in a manner
consistent with the requirements of this License, to extend the patent
license to downstream recipients. "Knowingly relying" means you have
actual knowledge that, but for the patent license, your conveying the
covered work in a country, or your recipient's use of the covered work
in a country, would infringe one or more identifiable patents in that
country that you have reason to believe are valid.
If, pursuant to or in connection with a single transaction or
arrangement, you convey, or propagate by procuring conveyance of, a
covered work, and grant a patent license to some of the parties
receiving the covered work authorizing them to use, propagate, modify
or convey a specific copy of the covered work, then the patent license
you grant is automatically extended to all recipients of the covered
work and works based on it.
A patent license is "discriminatory" if it does not include within
the scope of its coverage, prohibits the exercise of, or is
conditioned on the non-exercise of one or more of the rights that are
specifically granted under this License. You may not convey a covered
work if you are a party to an arrangement with a third party that is
in the business of distributing software, under which you make payment
to the third party based on the extent of your activity of conveying
the work, and under which the third party grants, to any of the
parties who would receive the covered work from you, a discriminatory
patent license (a) in connection with copies of the covered work
conveyed by you (or copies made from those copies), or (b) primarily
for and in connection with specific products or compilations that
contain the covered work, unless you entered into that arrangement,
or that patent license was granted, prior to 28 March 2007.
Nothing in this License shall be construed as excluding or limiting
any implied license or other defenses to infringement that may
otherwise be available to you under applicable patent law.
12. No Surrender of Others' Freedom.
If conditions are imposed on you (whether by court order, agreement or
otherwise) that contradict the conditions of this License, they do not
excuse you from the conditions of this License. If you cannot convey a
covered work so as to satisfy simultaneously your obligations under this
License and any other pertinent obligations, then as a consequence you may
not convey it at all. For example, if you agree to terms that obligate you
to collect a royalty for further conveying from those to whom you convey
the Program, the only way you could satisfy both those terms and this
License would be to refrain entirely from conveying the Program.
13. Use with the GNU Affero General Public License.
Notwithstanding any other provision of this License, you have
permission to link or combine any covered work with a work licensed
under version 3 of the GNU Affero General Public License into a single
combined work, and to convey the resulting work. The terms of this
License will continue to apply to the part which is the covered work,
but the special requirements of the GNU Affero General Public License,
section 13, concerning interaction through a network will apply to the
combination as such.
14. Revised Versions of this License.
The Free Software Foundation may publish revised and/or new versions of
the GNU General Public License from time to time. Such new versions will
be similar in spirit to the present version, but may differ in detail to
address new problems or concerns.
Each version is given a distinguishing version number. If the
Program specifies that a certain numbered version of the GNU General
Public License "or any later version" applies to it, you have the
option of following the terms and conditions either of that numbered
version or of any later version published by the Free Software
Foundation. If the Program does not specify a version number of the
GNU General Public License, you may choose any version ever published
by the Free Software Foundation.
If the Program specifies that a proxy can decide which future
versions of the GNU General Public License can be used, that proxy's
public statement of acceptance of a version permanently authorizes you
to choose that version for the Program.
Later license versions may give you additional or different
permissions. However, no additional obligations are imposed on any
author or copyright holder as a result of your choosing to follow a
later version.
15. Disclaimer of Warranty.
THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
16. Limitation of Liability.
IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
SUCH DAMAGES.
17. Interpretation of Sections 15 and 16.
If the disclaimer of warranty and limitation of liability provided
above cannot be given local legal effect according to their terms,
reviewing courts shall apply local law that most closely approximates
an absolute waiver of all civil liability in connection with the
Program, unless a warranty or assumption of liability accompanies a
copy of the Program in return for a fee.
END OF TERMS AND CONDITIONS
How to Apply These Terms to Your New Programs
If you develop a new program, and you want it to be of the greatest
possible use to the public, the best way to achieve this is to make it
free software which everyone can redistribute and change under these terms.
To do so, attach the following notices to the program. It is safest
to attach them to the start of each source file to most effectively
state the exclusion of warranty; and each file should have at least
the "copyright" line and a pointer to where the full notice is found.
<one line to give the program's name and a brief idea of what it does.>
Copyright (C) <year> <name of author>
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see <https://www.gnu.org/licenses/>.
Also add information on how to contact you by electronic and paper mail.
If the program does terminal interaction, make it output a short
notice like this when it starts in an interactive mode:
<program> Copyright (C) <year> <name of author>
This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
This is free software, and you are welcome to redistribute it
under certain conditions; type `show c' for details.
The hypothetical commands `show w' and `show c' should show the appropriate
parts of the General Public License. Of course, your program's commands
might be different; for a GUI interface, you would use an "about box".
You should also get your employer (if you work as a programmer) or school,
if any, to sign a "copyright disclaimer" for the program, if necessary.
For more information on this, and how to apply and follow the GNU GPL, see
<https://www.gnu.org/licenses/>.
The GNU General Public License does not permit incorporating your program
into proprietary programs. If your program is a subroutine library, you
may consider it more useful to permit linking proprietary applications with
the library. If this is what you want to do, use the GNU Lesser General
Public License instead of this License. But first, please read
<https://www.gnu.org/licenses/why-not-lgpl.html>.

39
annotator/clip/__init__.py
View File

@ -1,39 +0,0 @@
import torch
from transformers import CLIPProcessor, CLIPVisionModel
from modules import devices
import os
from annotator.annotator_path import clip_vision_path
remote_model_path = "https://huggingface.co/openai/clip-vit-large-patch14/resolve/main/pytorch_model.bin"
clip_path = clip_vision_path
print(f'ControlNet ClipVision location: {clip_path}')
clip_proc = None
clip_vision_model = None
def apply_clip(img):
global clip_proc, clip_vision_model
if clip_vision_model is None:
modelpath = os.path.join(clip_path, 'pytorch_model.bin')
if not os.path.exists(modelpath):
from basicsr.utils.download_util import load_file_from_url
load_file_from_url(remote_model_path, model_dir=clip_path)
clip_proc = CLIPProcessor.from_pretrained(clip_path)
clip_vision_model = CLIPVisionModel.from_pretrained(clip_path)
with torch.no_grad():
clip_vision_model = clip_vision_model.to(devices.get_device_for("controlnet"))
style_for_clip = clip_proc(images=img, return_tensors="pt")['pixel_values']
style_feat = clip_vision_model(style_for_clip.to(devices.get_device_for("controlnet")))['last_hidden_state']
return style_feat
def unload_clip_model():
global clip_proc, clip_vision_model
if clip_vision_model is not None:
clip_vision_model.cpu()

171
annotator/clip_vision/config.json
View File

@ -1,171 +0,0 @@
{
"_name_or_path": "clip-vit-large-patch14/",
"architectures": [
"CLIPModel"
],
"initializer_factor": 1.0,
"logit_scale_init_value": 2.6592,
"model_type": "clip",
"projection_dim": 768,
"text_config": {
"_name_or_path": "",
"add_cross_attention": false,
"architectures": null,
"attention_dropout": 0.0,
"bad_words_ids": null,
"bos_token_id": 0,
"chunk_size_feed_forward": 0,
"cross_attention_hidden_size": null,
"decoder_start_token_id": null,
"diversity_penalty": 0.0,
"do_sample": false,
"dropout": 0.0,
"early_stopping": false,
"encoder_no_repeat_ngram_size": 0,
"eos_token_id": 2,
"finetuning_task": null,
"forced_bos_token_id": null,
"forced_eos_token_id": null,
"hidden_act": "quick_gelu",
"hidden_size": 768,
"id2label": {
"0": "LABEL_0",
"1": "LABEL_1"
},
"initializer_factor": 1.0,
"initializer_range": 0.02,
"intermediate_size": 3072,
"is_decoder": false,
"is_encoder_decoder": false,
"label2id": {
"LABEL_0": 0,
"LABEL_1": 1
},
"layer_norm_eps": 1e-05,
"length_penalty": 1.0,
"max_length": 20,
"max_position_embeddings": 77,
"min_length": 0,
"model_type": "clip_text_model",
"no_repeat_ngram_size": 0,
"num_attention_heads": 12,
"num_beam_groups": 1,
"num_beams": 1,
"num_hidden_layers": 12,
"num_return_sequences": 1,
"output_attentions": false,
"output_hidden_states": false,
"output_scores": false,
"pad_token_id": 1,
"prefix": null,
"problem_type": null,
"projection_dim" : 768,
"pruned_heads": {},
"remove_invalid_values": false,
"repetition_penalty": 1.0,
"return_dict": true,
"return_dict_in_generate": false,
"sep_token_id": null,
"task_specific_params": null,
"temperature": 1.0,
"tie_encoder_decoder": false,
"tie_word_embeddings": true,
"tokenizer_class": null,
"top_k": 50,
"top_p": 1.0,
"torch_dtype": null,
"torchscript": false,
"transformers_version": "4.16.0.dev0",
"use_bfloat16": false,
"vocab_size": 49408
},
"text_config_dict": {
"hidden_size": 768,
"intermediate_size": 3072,
"num_attention_heads": 12,
"num_hidden_layers": 12,
"projection_dim": 768
},
"torch_dtype": "float32",
"transformers_version": null,
"vision_config": {
"_name_or_path": "",
"add_cross_attention": false,
"architectures": null,
"attention_dropout": 0.0,
"bad_words_ids": null,
"bos_token_id": null,
"chunk_size_feed_forward": 0,
"cross_attention_hidden_size": null,
"decoder_start_token_id": null,
"diversity_penalty": 0.0,
"do_sample": false,
"dropout": 0.0,
"early_stopping": false,
"encoder_no_repeat_ngram_size": 0,
"eos_token_id": null,
"finetuning_task": null,
"forced_bos_token_id": null,
"forced_eos_token_id": null,
"hidden_act": "quick_gelu",
"hidden_size": 1024,
"id2label": {
"0": "LABEL_0",
"1": "LABEL_1"
},
"image_size": 224,
"initializer_factor": 1.0,
"initializer_range": 0.02,
"intermediate_size": 4096,
"is_decoder": false,
"is_encoder_decoder": false,
"label2id": {
"LABEL_0": 0,
"LABEL_1": 1
},
"layer_norm_eps": 1e-05,
"length_penalty": 1.0,
"max_length": 20,
"min_length": 0,
"model_type": "clip_vision_model",
"no_repeat_ngram_size": 0,
"num_attention_heads": 16,
"num_beam_groups": 1,
"num_beams": 1,
"num_hidden_layers": 24,
"num_return_sequences": 1,
"output_attentions": false,
"output_hidden_states": false,
"output_scores": false,
"pad_token_id": null,
"patch_size": 14,
"prefix": null,
"problem_type": null,
"projection_dim" : 768,
"pruned_heads": {},
"remove_invalid_values": false,
"repetition_penalty": 1.0,
"return_dict": true,
"return_dict_in_generate": false,
"sep_token_id": null,
"task_specific_params": null,
"temperature": 1.0,
"tie_encoder_decoder": false,
"tie_word_embeddings": true,
"tokenizer_class": null,
"top_k": 50,
"top_p": 1.0,
"torch_dtype": null,
"torchscript": false,
"transformers_version": "4.16.0.dev0",
"use_bfloat16": false
},
"vision_config_dict": {
"hidden_size": 1024,
"intermediate_size": 4096,
"num_attention_heads": 16,
"num_hidden_layers": 24,
"patch_size": 14,
"projection_dim": 768
}
}

48895
annotator/clip_vision/merges.txt
View File

File diff suppressed because it is too large Load Diff

19
annotator/clip_vision/preprocessor_config.json
View File

@ -1,19 +0,0 @@
{
"crop_size": 224,
"do_center_crop": true,
"do_normalize": true,
"do_resize": true,
"feature_extractor_type": "CLIPFeatureExtractor",
"image_mean": [
0.48145466,
0.4578275,
0.40821073
],
"image_std": [
0.26862954,
0.26130258,
0.27577711
],
"resample": 3,
"size": 224
}

98393
annotator/clip_vision/tokenizer.json
View File

File diff suppressed because it is too large Load Diff

34
annotator/clip_vision/tokenizer_config.json
View File

@ -1,34 +0,0 @@
{
"unk_token": {
"content": "<|endoftext|>",
"single_word": false,
"lstrip": false,
"rstrip": false,
"normalized": true,
"__type": "AddedToken"
},
"bos_token": {
"content": "<|startoftext|>",
"single_word": false,
"lstrip": false,
"rstrip": false,
"normalized": true,
"__type": "AddedToken"
},
"eos_token": {
"content": "<|endoftext|>",
"single_word": false,
"lstrip": false,
"rstrip": false,
"normalized": true,
"__type": "AddedToken"
},
"pad_token": "<|endoftext|>",
"add_prefix_space": false,
"errors": "replace",
"do_lower_case": true,
"name_or_path": "openai/clip-vit-base-patch32",
"model_max_length": 77,
"special_tokens_map_file": "./special_tokens_map.json",
"tokenizer_class": "CLIPTokenizer"
}

1
annotator/clip_vision/vocab.json
View File

File diff suppressed because one or more lines are too long

123
annotator/clipvision/__init__.py Normal file
View File

@ -0,0 +1,123 @@
import os
import torch
from modules import devices
from modules.modelloader import load_file_from_url
from annotator.annotator_path import models_path
from transformers import CLIPVisionModelWithProjection, CLIPVisionConfig, CLIPImageProcessor, modeling_utils
config_clip_g = {
"attention_dropout": 0.0,
"dropout": 0.0,
"hidden_act": "gelu",
"hidden_size": 1664,
"image_size": 224,
"initializer_factor": 1.0,
"initializer_range": 0.02,
"intermediate_size": 8192,
"layer_norm_eps": 1e-05,
"model_type": "clip_vision_model",
"num_attention_heads": 16,
"num_channels": 3,
"num_hidden_layers": 48,
"patch_size": 14,
"projection_dim": 1280,
"torch_dtype": "float32"
}
config_clip_h = {
"attention_dropout": 0.0,
"dropout": 0.0,
"hidden_act": "gelu",
"hidden_size": 1280,
"image_size": 224,
"initializer_factor": 1.0,
"initializer_range": 0.02,
"intermediate_size": 5120,
"layer_norm_eps": 1e-05,
"model_type": "clip_vision_model",
"num_attention_heads": 16,
"num_channels": 3,
"num_hidden_layers": 32,
"patch_size": 14,
"projection_dim": 1024,
"torch_dtype": "float32"
}
config_clip_vitl = {
"attention_dropout": 0.0,
"dropout": 0.0,
"hidden_act": "quick_gelu",
"hidden_size": 1024,
"image_size": 224,
"initializer_factor": 1.0,
"initializer_range": 0.02,
"intermediate_size": 4096,
"layer_norm_eps": 1e-05,
"model_type": "clip_vision_model",
"num_attention_heads": 16,
"num_channels": 3,
"num_hidden_layers": 24,
"patch_size": 14,
"projection_dim": 768,
"torch_dtype": "float32"
}
configs = {
'clip_g': config_clip_g,
'clip_h': config_clip_h,
'clip_vitl': config_clip_vitl,
}
downloads = {
'clip_vitl': 'https://huggingface.co/openai/clip-vit-large-patch14/resolve/main/pytorch_model.bin',
'clip_g': 'https://huggingface.co/lllyasviel/Annotators/resolve/main/clip_g.pth',
'clip_h': 'https://huggingface.co/h94/IP-Adapter/resolve/main/models/image_encoder/pytorch_model.bin'
}
class ClipVisionDetector:
def __init__(self, config):
assert config in downloads
self.download_link = downloads[config]
self.model_path = os.path.join(models_path, 'clip_vision')
self.file_name = config + '.pth'
self.config = configs[config]
self.device = devices.get_device_for("controlnet")
os.makedirs(self.model_path, exist_ok=True)
file_path = os.path.join(self.model_path, self.file_name)
if not os.path.exists(file_path):
load_file_from_url(url=self.download_link, model_dir=self.model_path, file_name=self.file_name)
config = CLIPVisionConfig(**self.config)
self.model = CLIPVisionModelWithProjection(config)
self.processor = CLIPImageProcessor(crop_size=224,
do_center_crop=True,
do_convert_rgb=True,
do_normalize=True,
do_resize=True,
image_mean=[0.48145466, 0.4578275, 0.40821073],
image_std=[0.26862954, 0.26130258, 0.27577711],
resample=3,
size=224)
sd = torch.load(file_path, map_location=torch.device('cpu'))
self.model.load_state_dict(sd, strict=False)
del sd
self.model.eval()
self.model.cpu()
def unload_model(self):
if self.model is not None:
self.model.to('meta')
def __call__(self, input_image):
with torch.no_grad():
clip_vision_model = self.model.cpu()
feat = self.processor(images=input_image, return_tensors="pt")
feat['pixel_values'] = feat['pixel_values'].cpu()
result = clip_vision_model(**feat, output_hidden_states=True)
result['hidden_states'] = [v.to(devices.get_device_for("controlnet")) for v in result['hidden_states']]
result = {k: v.to(devices.get_device_for("controlnet")) if isinstance(v, torch.Tensor) else v for k, v in result.items()}
return result

58
annotator/shuffle/__init__.py
View File

@ -1,8 +1,6 @@
import random
import cv2
import numpy as np
from annotator.util import make_noise_disk, img2mask
from annotator.util import make_noise_disk
class ContentShuffleDetector:
@ -18,57 +16,3 @@ class ContentShuffleDetector:
y = make_noise_disk(h, w, 1, f) * float(H - 1)
flow = np.concatenate([x, y], axis=2).astype(np.float32)
return cv2.remap(img, flow, None, cv2.INTER_LINEAR)
class ColorShuffleDetector:
def __call__(self, img):
H, W, C = img.shape
F = np.random.randint(64, 384)
A = make_noise_disk(H, W, 3, F)
B = make_noise_disk(H, W, 3, F)
C = (A + B) / 2.0
A = (C + (A - C) * 3.0).clip(0, 1)
B = (C + (B - C) * 3.0).clip(0, 1)
L = img.astype(np.float32) / 255.0
Y = A * L + B * (1 - L)
Y -= np.min(Y, axis=(0, 1), keepdims=True)
Y /= np.maximum(np.max(Y, axis=(0, 1), keepdims=True), 1e-5)
Y *= 255.0
return Y.clip(0, 255).astype(np.uint8)
class GrayDetector:
def __call__(self, img):
eps = 1e-5
X = img.astype(np.float32)
r, g, b = X[:, :, 0], X[:, :, 1], X[:, :, 2]
kr, kg, kb = [random.random() + eps for _ in range(3)]
ks = kr + kg + kb
kr /= ks
kg /= ks
kb /= ks
Y = r * kr + g * kg + b * kb
Y = np.stack([Y] * 3, axis=2)
return Y.clip(0, 255).astype(np.uint8)
class DownSampleDetector:
def __call__(self, img, level=3, k=16.0):
h = img.astype(np.float32)
for _ in range(level):
h += np.random.normal(loc=0.0, scale=k, size=h.shape)
h = cv2.pyrDown(h)
for _ in range(level):
h = cv2.pyrUp(h)
h += np.random.normal(loc=0.0, scale=k, size=h.shape)
return h.clip(0, 255).astype(np.uint8)
class Image2MaskShuffleDetector:
def __init__(self, resolution=(640, 512)):
self.H, self.W = resolution
def __call__(self, img):
m = img2mask(img, self.H, self.W)
m *= 255.0
return m.clip(0, 255).astype(np.uint8)

17
annotator/util.py
View File

@ -60,20 +60,3 @@ def safe_step(x, step=2):
y = x.astype(np.float32) * float(step + 1)
y = y.astype(np.int32).astype(np.float32) / float(step)
return y
def img2mask(img, H, W, low=10, high=90):
assert img.ndim == 3 or img.ndim == 2
assert img.dtype == np.uint8
if img.ndim == 3:
y = img[:, :, random.randrange(0, img.shape[2])]
else:
y = img
y = cv2.resize(y, (W, H), interpolation=cv2.INTER_CUBIC)
if random.uniform(0, 1) < 0.5:
y = 255 - y
return y < np.percentile(y, random.randrange(low, high))

4
internal_controlnet/external_code.py
View File

@ -283,7 +283,7 @@ def get_max_models_num():
Fetch the maximum number of allowed ControlNet models.
"""
max_models_num = shared.opts.data.get("control_net_max_models_num", 1)
max_models_num = shared.opts.data.get("control_net_unit_count", 3)
return max_models_num
@ -362,7 +362,7 @@ def update_cn_script_in_place(
return
# fill in remaining parameters to satisfy max models, just in case script needs it.
max_models = shared.opts.data.get("control_net_max_models_num", 1)
max_models = shared.opts.data.get("control_net_unit_count", 3)
cn_units = cn_units + [ControlNetUnit(enabled=False)] * max(max_models - len(cn_units), 0)
cn_script_args_diff = 0

79
models/cldm_v15.yaml
View File

@ -1,79 +0,0 @@
model:
target: cldm.cldm.ControlLDM
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
control_key: "hint"
image_size: 64
channels: 4
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
only_mid_control: False
control_stage_config:
target: cldm.cldm.ControlNet
params:
image_size: 32 # unused
in_channels: 4
hint_channels: 3
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
unet_config:
target: cldm.cldm.ControlledUnetModel
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

85
models/cldm_v21.yaml
View File

@ -1,85 +0,0 @@
model:
target: cldm.cldm.ControlLDM
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
control_key: "hint"
image_size: 64
channels: 4
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
only_mid_control: False
control_stage_config:
target: cldm.cldm.ControlNet
params:
use_checkpoint: True
image_size: 32 # unused
in_channels: 4
hint_channels: 3
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_head_channels: 64 # need to fix for flash-attn
use_spatial_transformer: True
use_linear_in_transformer: True
transformer_depth: 1
context_dim: 1024
legacy: False
unet_config:
target: cldm.cldm.ControlledUnetModel
params:
use_checkpoint: True
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_head_channels: 64 # need to fix for flash-attn
use_spatial_transformer: True
use_linear_in_transformer: True
transformer_depth: 1
context_dim: 1024
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
#attn_type: "vanilla-xformers"
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenOpenCLIPEmbedder
params:
freeze: True
layer: "penultimate"

79
models/control_sd15_canny.yaml
View File

@ -1,79 +0,0 @@
model:
target: cldm.cldm.ControlLDM
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
control_key: "hint"
image_size: 64
channels: 4
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
only_mid_control: False
control_stage_config:
target: cldm.cldm.ControlNet
params:
image_size: 32 # unused
in_channels: 4
hint_channels: 3
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
unet_config:
target: cldm.cldm.ControlledUnetModel
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

79
models/control_sd15_depth.yaml
View File

@ -1,79 +0,0 @@
model:
target: cldm.cldm.ControlLDM
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
control_key: "hint"
image_size: 64
channels: 4
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
only_mid_control: False
control_stage_config:
target: cldm.cldm.ControlNet
params:
image_size: 32 # unused
in_channels: 4
hint_channels: 3
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
unet_config:
target: cldm.cldm.ControlledUnetModel
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

79
models/control_sd15_hed.yaml
View File

@ -1,79 +0,0 @@
model:
target: cldm.cldm.ControlLDM
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
control_key: "hint"
image_size: 64
channels: 4
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
only_mid_control: False
control_stage_config:
target: cldm.cldm.ControlNet
params:
image_size: 32 # unused
in_channels: 4
hint_channels: 3
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
unet_config:
target: cldm.cldm.ControlledUnetModel
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

79
models/control_sd15_mlsd.yaml
View File

@ -1,79 +0,0 @@
model:
target: cldm.cldm.ControlLDM
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
control_key: "hint"
image_size: 64
channels: 4
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
only_mid_control: False
control_stage_config:
target: cldm.cldm.ControlNet
params:
image_size: 32 # unused
in_channels: 4
hint_channels: 3
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
unet_config:
target: cldm.cldm.ControlledUnetModel
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

79
models/control_sd15_normal.yaml
View File

@ -1,79 +0,0 @@
model:
target: cldm.cldm.ControlLDM
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
control_key: "hint"
image_size: 64
channels: 4
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
only_mid_control: False
control_stage_config:
target: cldm.cldm.ControlNet
params:
image_size: 32 # unused
in_channels: 4
hint_channels: 3
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
unet_config:
target: cldm.cldm.ControlledUnetModel
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

79
models/control_sd15_openpose.yaml
View File

@ -1,79 +0,0 @@
model:
target: cldm.cldm.ControlLDM
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
control_key: "hint"
image_size: 64
channels: 4
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
only_mid_control: False
control_stage_config:
target: cldm.cldm.ControlNet
params:
image_size: 32 # unused
in_channels: 4
hint_channels: 3
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
unet_config:
target: cldm.cldm.ControlledUnetModel
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

79
models/control_sd15_scribble.yaml
View File

@ -1,79 +0,0 @@
model:
target: cldm.cldm.ControlLDM
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
control_key: "hint"
image_size: 64
channels: 4
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
only_mid_control: False
control_stage_config:
target: cldm.cldm.ControlNet
params:
image_size: 32 # unused
in_channels: 4
hint_channels: 3
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
unet_config:
target: cldm.cldm.ControlledUnetModel
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

79
models/control_sd15_seg.yaml
View File

@ -1,79 +0,0 @@
model:
target: cldm.cldm.ControlLDM
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
control_key: "hint"
image_size: 64
channels: 4
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
only_mid_control: False
control_stage_config:
target: cldm.cldm.ControlNet
params:
image_size: 32 # unused
in_channels: 4
hint_channels: 3
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
unet_config:
target: cldm.cldm.ControlledUnetModel
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

79
models/control_v11e_sd15_ip2p.yaml
View File

@ -1,79 +0,0 @@
model:
target: cldm.cldm.ControlLDM
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
control_key: "hint"
image_size: 64
channels: 4
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
only_mid_control: False
control_stage_config:
target: cldm.cldm.ControlNet
params:
image_size: 32 # unused
in_channels: 4
hint_channels: 3
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
unet_config:
target: cldm.cldm.ControlledUnetModel
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

80
models/control_v11e_sd15_shuffle.yaml
View File

@ -1,80 +0,0 @@
model:
target: cldm.cldm.ControlLDM
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
control_key: "hint"
image_size: 64
channels: 4
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
only_mid_control: False
global_average_pooling: True
control_stage_config:
target: cldm.cldm.ControlNet
params:
image_size: 32 # unused
in_channels: 4
hint_channels: 3
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
unet_config:
target: cldm.cldm.ControlledUnetModel
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

79
models/control_v11f1e_sd15_tile.yaml
View File

@ -1,79 +0,0 @@
model:
target: cldm.cldm.ControlLDM
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
control_key: "hint"
image_size: 64
channels: 4
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
only_mid_control: False
control_stage_config:
target: cldm.cldm.ControlNet
params:
image_size: 32 # unused
in_channels: 4
hint_channels: 3
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
unet_config:
target: cldm.cldm.ControlledUnetModel
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

79
models/control_v11f1p_sd15_depth.yaml
View File

@ -1,79 +0,0 @@
model:
target: cldm.cldm.ControlLDM
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
control_key: "hint"
image_size: 64
channels: 4
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
only_mid_control: False
control_stage_config:
target: cldm.cldm.ControlNet
params:
image_size: 32 # unused
in_channels: 4
hint_channels: 3
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
unet_config:
target: cldm.cldm.ControlledUnetModel
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

79
models/control_v11p_sd15_canny.yaml
View File

@ -1,79 +0,0 @@
model:
target: cldm.cldm.ControlLDM
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
control_key: "hint"
image_size: 64
channels: 4
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
only_mid_control: False
control_stage_config:
target: cldm.cldm.ControlNet
params:
image_size: 32 # unused
in_channels: 4
hint_channels: 3
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
unet_config:
target: cldm.cldm.ControlledUnetModel
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

79
models/control_v11p_sd15_inpaint.yaml
View File

@ -1,79 +0,0 @@
model:
target: cldm.cldm.ControlLDM
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
control_key: "hint"
image_size: 64
channels: 4
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
only_mid_control: False
control_stage_config:
target: cldm.cldm.ControlNet
params:
image_size: 32 # unused
in_channels: 4
hint_channels: 3
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
unet_config:
target: cldm.cldm.ControlledUnetModel
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

79
models/control_v11p_sd15_lineart.yaml
View File

@ -1,79 +0,0 @@
model:
target: cldm.cldm.ControlLDM
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
control_key: "hint"
image_size: 64
channels: 4
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
only_mid_control: False
control_stage_config:
target: cldm.cldm.ControlNet
params:
image_size: 32 # unused
in_channels: 4
hint_channels: 3
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
unet_config:
target: cldm.cldm.ControlledUnetModel
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

79
models/control_v11p_sd15_mlsd.yaml
View File

@ -1,79 +0,0 @@
model:
target: cldm.cldm.ControlLDM
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
control_key: "hint"
image_size: 64
channels: 4
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
only_mid_control: False
control_stage_config:
target: cldm.cldm.ControlNet
params:
image_size: 32 # unused
in_channels: 4
hint_channels: 3
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
unet_config:
target: cldm.cldm.ControlledUnetModel
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

79
models/control_v11p_sd15_normalbae.yaml
View File

@ -1,79 +0,0 @@
model:
target: cldm.cldm.ControlLDM
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
control_key: "hint"
image_size: 64
channels: 4
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
only_mid_control: False
control_stage_config:
target: cldm.cldm.ControlNet
params:
image_size: 32 # unused
in_channels: 4
hint_channels: 3
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
unet_config:
target: cldm.cldm.ControlledUnetModel
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

79
models/control_v11p_sd15_openpose.yaml
View File

@ -1,79 +0,0 @@
model:
target: cldm.cldm.ControlLDM
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
control_key: "hint"
image_size: 64
channels: 4
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
only_mid_control: False
control_stage_config:
target: cldm.cldm.ControlNet
params:
image_size: 32 # unused
in_channels: 4
hint_channels: 3
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
unet_config:
target: cldm.cldm.ControlledUnetModel
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

79
models/control_v11p_sd15_scribble.yaml
View File

@ -1,79 +0,0 @@
model:
target: cldm.cldm.ControlLDM
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
control_key: "hint"
image_size: 64
channels: 4
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
only_mid_control: False
control_stage_config:
target: cldm.cldm.ControlNet
params:
image_size: 32 # unused
in_channels: 4
hint_channels: 3
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
unet_config:
target: cldm.cldm.ControlledUnetModel
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

79
models/control_v11p_sd15_seg.yaml
View File

@ -1,79 +0,0 @@
model:
target: cldm.cldm.ControlLDM
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
control_key: "hint"
image_size: 64
channels: 4
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
only_mid_control: False
control_stage_config:
target: cldm.cldm.ControlNet
params:
image_size: 32 # unused
in_channels: 4
hint_channels: 3
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
unet_config:
target: cldm.cldm.ControlledUnetModel
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

79
models/control_v11p_sd15_softedge.yaml
View File

@ -1,79 +0,0 @@
model:
target: cldm.cldm.ControlLDM
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
control_key: "hint"
image_size: 64
channels: 4
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
only_mid_control: False
control_stage_config:
target: cldm.cldm.ControlNet
params:
image_size: 32 # unused
in_channels: 4
hint_channels: 3
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
unet_config:
target: cldm.cldm.ControlledUnetModel
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

79
models/control_v11p_sd15s2_lineart_anime.yaml
View File

@ -1,79 +0,0 @@
model:
target: cldm.cldm.ControlLDM
params:
linear_start: 0.00085
linear_end: 0.0120
num_timesteps_cond: 1
log_every_t: 200
timesteps: 1000
first_stage_key: "jpg"
cond_stage_key: "txt"
control_key: "hint"
image_size: 64
channels: 4
cond_stage_trainable: false
conditioning_key: crossattn
monitor: val/loss_simple_ema
scale_factor: 0.18215
use_ema: False
only_mid_control: False
control_stage_config:
target: cldm.cldm.ControlNet
params:
image_size: 32 # unused
in_channels: 4
hint_channels: 3
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
unet_config:
target: cldm.cldm.ControlledUnetModel
params:
image_size: 32 # unused
in_channels: 4
out_channels: 4
model_channels: 320
attention_resolutions: [ 4, 2, 1 ]
num_res_blocks: 2
channel_mult: [ 1, 2, 4, 4 ]
num_heads: 8
use_spatial_transformer: True
transformer_depth: 1
context_dim: 768
use_checkpoint: True
legacy: False
first_stage_config:
target: ldm.models.autoencoder.AutoencoderKL
params:
embed_dim: 4
monitor: val/rec_loss
ddconfig:
double_z: true
z_channels: 4
resolution: 256
in_channels: 3
out_ch: 3
ch: 128
ch_mult:
- 1
- 2
- 4
- 4
num_res_blocks: 2
attn_resolutions: []
dropout: 0.0
lossconfig:
target: torch.nn.Identity
cond_stage_config:
target: ldm.modules.encoders.modules.FrozenCLIPEmbedder

9
models/image_adapter_v14.yaml
View File

@ -1,9 +0,0 @@
model:
target: tencentarc.t21_adapter
params:
channels: [320, 640, 1280, 1280]
nums_rb: 2
ksize: 1
sk: true
cin: 192
use_conv: false

1
models/put_controlnet_models_here.txt Normal file
View File

@ -0,0 +1 @@
put_controlnet_models_here

9
models/sketch_adapter_v14.yaml
View File

@ -1,9 +0,0 @@
model:
target: tencentarc.t21_adapter
params:
channels: [320, 640, 1280, 1280]
nums_rb: 2
ksize: 1
sk: true
cin: 64
use_conv: false

9
models/t2iadapter_canny_sd14v1.yaml
View File

@ -1,9 +0,0 @@
model:
target: tencentarc.t21_adapter
params:
channels: [320, 640, 1280, 1280]
nums_rb: 2
ksize: 1
sk: true
cin: 64
use_conv: false

9
models/t2iadapter_canny_sd15v2.yaml
View File

@ -1,9 +0,0 @@
model:
target: tencentarc.t21_adapter
params:
channels: [320, 640, 1280, 1280]
nums_rb: 2
ksize: 1
sk: true
cin: 64
use_conv: false

6
models/t2iadapter_color_sd14v1.yaml
View File

@ -1,6 +0,0 @@
model:
target: scripts.adapter.Adapter_light
params:
channels: [320, 640, 1280, 1280]
nums_rb: 4
cin: 192

9
models/t2iadapter_depth_sd14v1.yaml
View File

@ -1,9 +0,0 @@
model:
target: tencentarc.t21_adapter
params:
channels: [320, 640, 1280, 1280]
nums_rb: 2
ksize: 1
sk: true
cin: 192
use_conv: false

9
models/t2iadapter_depth_sd15v2.yaml
View File

@ -1,9 +0,0 @@
model:
target: tencentarc.t21_adapter
params:
channels: [320, 640, 1280, 1280]
nums_rb: 2
ksize: 1
sk: true
cin: 192
use_conv: false

9
models/t2iadapter_keypose_sd14v1.yaml
View File

@ -1,9 +0,0 @@
model:
target: tencentarc.t21_adapter
params:
channels: [320, 640, 1280, 1280]
nums_rb: 2
ksize: 1
sk: true
cin: 192
use_conv: false

9
models/t2iadapter_openpose_sd14v1.yaml
View File

@ -1,9 +0,0 @@
model:
target: tencentarc.t21_adapter
params:
channels: [320, 640, 1280, 1280]
nums_rb: 2
ksize: 1
sk: true
cin: 192
use_conv: false

9
models/t2iadapter_seg_sd14v1.yaml
View File

@ -1,9 +0,0 @@
model:
target: tencentarc.t21_adapter
params:
channels: [320, 640, 1280, 1280]
nums_rb: 2
ksize: 1
sk: true
cin: 192
use_conv: false

9
models/t2iadapter_sketch_sd14v1.yaml
View File

@ -1,9 +0,0 @@
model:
target: tencentarc.t21_adapter
params:
channels: [320, 640, 1280, 1280]
nums_rb: 2
ksize: 1
sk: true
cin: 64
use_conv: false

9
models/t2iadapter_sketch_sd15v2.yaml
View File

@ -1,9 +0,0 @@
model:
target: tencentarc.t21_adapter
params:
channels: [320, 640, 1280, 1280]
nums_rb: 2
ksize: 1
sk: true
cin: 64
use_conv: false

8
models/t2iadapter_style_sd14v1.yaml
View File

@ -1,8 +0,0 @@
model:
target: scripts.adapter.StyleAdapter
params:
width: 1024
context_dim: 768
num_head: 8
n_layes: 3
num_token: 8

9
models/t2iadapter_zoedepth_sd15v1.yaml
View File

@ -1,9 +0,0 @@
model:
target: tencentarc.t21_adapter
params:
channels: [320, 640, 1280, 1280]
nums_rb: 2
ksize: 1
sk: true
cin: 192
use_conv: false

120
scripts/adapter.py
View File

@ -1,16 +1,11 @@
import torch
import torch.nn as nn
import importlib
from collections import OrderedDict
from omegaconf import OmegaConf
from copy import deepcopy
from modules import devices, lowvram, shared, scripts
cond_cast_unet = getattr(devices, 'cond_cast_unet', lambda x: x)
from ldm.modules.diffusionmodules.util import timestep_embedding
from ldm.modules.diffusionmodules.openaimodel import UNetModel
class TorchHijackForUnet:
@ -50,42 +45,12 @@ def align(hint, size):
return hint
def get_node_name(name, parent_name):
if len(name) <= len(parent_name):
return False, ''
p = name[:len(parent_name)]
if p != parent_name:
return False, ''
return True, name[len(parent_name):]
def get_obj_from_str(string, reload=False):
module, cls = string.rsplit(".", 1)
if reload:
module_imp = importlib.import_module(module)
importlib.reload(module_imp)
return getattr(importlib.import_module(module, package=None), cls)
class PlugableAdapter(nn.Module):
def __init__(self, state_dict, config_path, lowvram=False, base_model=None) -> None:
def __init__(self, control_model) -> None:
super().__init__()
self.config = OmegaConf.load(config_path)
model = Adapter
try:
self.target = self.config.model.target
model = get_obj_from_str(self.config.model.target)
except ImportError:
pass
self.control_model = model(**self.config.model.params)
self.control_model.load_state_dict(state_dict)
self.lowvram = lowvram
self.control_model = control_model
self.control = None
self.hint_cond = None
if not self.lowvram:
self.control_model.to(devices.get_device_for("controlnet"))
def reset(self):
self.control = None
@ -98,12 +63,21 @@ class PlugableAdapter(nn.Module):
self.hint_cond = cond_cast_unet(hint)
hint_in = cond_cast_unet(hint)
if hasattr(self.control_model, 'conv_in') and self.control_model.conv_in.in_channels == 64:
if hasattr(self.control_model, 'conv_in') and \
(self.control_model.conv_in.in_channels == 64 or self.control_model.conv_in.in_channels == 256):
hint_in = hint_in[:, 0:1, :, :]
self.control = self.control_model(hint_in)
return deepcopy(self.control)
def aggressive_lowvram(self):
self.to(devices.get_device_for("controlnet"))
return
def fullvram(self):
self.to(devices.get_device_for("controlnet"))
return
def conv_nd(dims, *args, **kwargs):
"""
@ -236,34 +210,76 @@ class ResnetBlock(nn.Module):
class Adapter(nn.Module):
def __init__(self, channels=[320, 640, 1280, 1280], nums_rb=3, cin=64, ksize=3, sk=False, use_conv=True):
def __init__(self, channels=[320, 640, 1280, 1280], nums_rb=3, cin=64, ksize=3, sk=False, use_conv=True, is_sdxl=True):
super(Adapter, self).__init__()
self.unshuffle = nn.PixelUnshuffle(8)
if is_sdxl:
self.pixel_shuffle = 16
downsample_avoided = [1]
downsample_layers = [2]
else:
self.pixel_shuffle = 8
downsample_avoided = []
downsample_layers = [3, 2, 1]
self.input_channels = cin // (self.pixel_shuffle * self.pixel_shuffle)
self.channels = channels
self.nums_rb = nums_rb
self.body = []
self.unshuffle = nn.PixelUnshuffle(self.pixel_shuffle)
for i in range(len(channels)):
for j in range(nums_rb):
if (i!=0) and (j==0):
self.body.append(ResnetBlock(channels[i-1], channels[i], down=True, ksize=ksize, sk=sk, use_conv=use_conv))
else:
self.body.append(ResnetBlock(channels[i], channels[i], down=False, ksize=ksize, sk=sk, use_conv=use_conv))
for r in range(nums_rb):
if i in downsample_layers and r == 0:
self.body.append(ResnetBlock(
channels[i - 1],
channels[i],
down=True,
ksize=ksize,
sk=sk,
use_conv=use_conv))
continue
if i in downsample_avoided and r == 0:
self.body.append(ResnetBlock(
channels[i - 1],
channels[i],
down=False,
ksize=ksize,
sk=sk,
use_conv=use_conv))
continue
self.body.append(ResnetBlock(
channels[i],
channels[i],
down=False,
ksize=ksize,
sk=sk,
use_conv=use_conv
))
self.body = nn.ModuleList(self.body)
self.conv_in = nn.Conv2d(cin, channels[0], 3, 1, 1)
def forward(self, x):
# unshuffle
self.to(x.device)
x = self.unshuffle(x)
# extract features
features = []
hs = []
x = self.conv_in(x)
for i in range(len(self.channels)):
for j in range(self.nums_rb):
idx = i*self.nums_rb +j
for r in range(self.nums_rb):
idx = i * self.nums_rb + r
x = self.body[idx](x)
features.append(x)
hs.append(x)
self.to('cpu')
return hs
return features
class LayerNorm(nn.LayerNorm):
"""Subclass torch's LayerNorm to handle fp16."""

2
scripts/api.py
View File

@ -77,7 +77,7 @@ def controlnet_api(_: gr.Blocks, app: FastAPI):
@app.get("/controlnet/settings")
async def settings():
max_models_num = external_code.get_max_models_num()
return {"control_net_max_models_num": max_models_num}
return {"control_net_unit_count": max_models_num}
cached_cn_preprocessors = global_state.cache_preprocessors(
global_state.cn_preprocessor_modules

310
scripts/cldm.py
View File

@ -1,108 +1,67 @@
import torch
import torch.nn as nn
from omegaconf import OmegaConf
from modules import devices, shared
cond_cast_unet = getattr(devices, 'cond_cast_unet', lambda x: x)
from ldm.util import exists
from ldm.modules.attention import SpatialTransformer
from ldm.modules.diffusionmodules.util import conv_nd, linear, zero_module, timestep_embedding
from ldm.modules.diffusionmodules.openaimodel import UNetModel, TimestepEmbedSequential, ResBlock, Downsample, AttentionBlock
from modules import devices
class TorchHijackForUnet:
"""
This is torch, but with cat that resizes tensors to appropriate dimensions if they do not match;
this makes it possible to create pictures with dimensions that are multiples of 8 rather than 64
"""
def __getattr__(self, item):
if item == 'cat':
return self.cat
if hasattr(torch, item):
return getattr(torch, item)
raise AttributeError("'{}' object has no attribute '{}'".format(type(self).__name__, item))
def cat(self, tensors, *args, **kwargs):
if len(tensors) == 2:
a, b = tensors
if a.shape[-2:] != b.shape[-2:]:
a = torch.nn.functional.interpolate(a, b.shape[-2:], mode="nearest")
tensors = (a, b)
return torch.cat(tensors, *args, **kwargs)
th = TorchHijackForUnet()
def align(hint, size):
b, c, h1, w1 = hint.shape
h, w = size
if h != h1 or w != w1:
hint = th.nn.functional.interpolate(hint, size=size, mode="nearest")
return hint
def get_node_name(name, parent_name):
if len(name) <= len(parent_name):
return False, ''
p = name[:len(parent_name)]
if p != parent_name:
return False, ''
return True, name[len(parent_name):]
try:
from sgm.modules.diffusionmodules.openaimodel import conv_nd, linear, zero_module, timestep_embedding, \
TimestepEmbedSequential, ResBlock, Downsample, SpatialTransformer, exists
using_sgm = True
except:
from ldm.modules.diffusionmodules.openaimodel import conv_nd, linear, zero_module, timestep_embedding, \
TimestepEmbedSequential, ResBlock, Downsample, SpatialTransformer, exists
using_sgm = False
class PlugableControlModel(nn.Module):
def __init__(self, state_dict, config_path, lowvram=False, base_model=None) -> None:
def __init__(self, config, state_dict=None):
super().__init__()
self.config = OmegaConf.load(config_path)
self.control_model = ControlNet(**self.config.model.params.control_stage_config.params)
if any([k.startswith("control_model.") for k, v in state_dict.items()]):
if 'difference' in state_dict and base_model is not None:
print('We will stop supporting diff models soon because of its lack of robustness.')
print('Please begin to use official models as soon as possible.')
self.config = config
self.control_model = ControlNet(**self.config).cpu()
if state_dict is not None:
self.control_model.load_state_dict(state_dict, strict=False)
self.gpu_component = None
self.is_control_lora = False
unet_state_dict = base_model.state_dict()
unet_state_dict_keys = unet_state_dict.keys()
final_state_dict = {}
counter = 0
for key in state_dict.keys():
if not key.startswith("control_model."):
continue
p = state_dict[key]
is_control, node_name = get_node_name(key, 'control_')
key_name = node_name.replace("model.", "") if is_control else key
if key_name in unet_state_dict_keys:
p_new = p + unet_state_dict[key_name].clone().cpu()
counter += 1
else:
p_new = p
final_state_dict[key] = p_new
print(f'Diff model cloned: {counter} values')
state_dict = final_state_dict
state_dict = {k.replace("control_model.", ""): v for k, v in state_dict.items() if k.startswith("control_model.")}
self.control_model.load_state_dict(state_dict)
if not lowvram:
self.control_model.to(devices.get_device_for("controlnet"))
def reset(self):
pass
def forward(self, *args, **kwargs):
return self.control_model(*args, **kwargs)
def aggressive_lowvram(self):
self.to('cpu')
def send_me_to_gpu(module, _):
if self.gpu_component == module:
return
if self.gpu_component is not None:
self.gpu_component.to('cpu')
module.to(devices.get_device_for("controlnet"))
self.gpu_component = module
self.control_model.time_embed.register_forward_pre_hook(send_me_to_gpu)
self.control_model.input_hint_block.register_forward_pre_hook(send_me_to_gpu)
self.control_model.label_emb.register_forward_pre_hook(send_me_to_gpu)
for m in self.control_model.input_blocks:
m.register_forward_pre_hook(send_me_to_gpu)
for m in self.control_model.zero_convs:
m.register_forward_pre_hook(send_me_to_gpu)
self.control_model.middle_block.register_forward_pre_hook(send_me_to_gpu)
self.control_model.middle_block_out.register_forward_pre_hook(send_me_to_gpu)
return
def fullvram(self):
self.to(devices.get_device_for("controlnet"))
return
class ControlNet(nn.Module):
def __init__(
self,
image_size,
in_channels,
model_channels,
hint_channels,
@ -112,75 +71,54 @@ class ControlNet(nn.Module):
channel_mult=(1, 2, 4, 8),
conv_resample=True,
dims=2,
num_classes=None,
use_checkpoint=False,
use_fp16=False,
use_fp16=True,
num_heads=-1,
num_head_channels=-1,
num_heads_upsample=-1,
use_scale_shift_norm=False,
resblock_updown=False,
use_new_attention_order=False,
use_spatial_transformer=False, # custom transformer support
transformer_depth=1, # custom transformer support
context_dim=None, # custom transformer support
# custom support for prediction of discrete ids into codebook of first stage vq model
use_spatial_transformer=True,
transformer_depth=1,
context_dim=None,
n_embed=None,
legacy=True,
legacy=False,
disable_self_attentions=None,
num_attention_blocks=None,
disable_middle_self_attn=False,
use_linear_in_transformer=False,
adm_in_channels=None,
transformer_depth_middle=None,
device=None,
global_average_pooling=False,
):
use_fp16 = getattr(devices, 'dtype_unet', devices.dtype) == th.float16 and not getattr(shared.cmd_opts, "no_half_controlnet", False)
super().__init__()
if use_spatial_transformer:
assert context_dim is not None, 'Fool!! You forgot to include the dimension of your cross-attention conditioning...'
if context_dim is not None:
assert use_spatial_transformer, 'Fool!! You forgot to use the spatial transformer for your cross-attention conditioning...'
from omegaconf.listconfig import ListConfig
if type(context_dim) == ListConfig:
context_dim = list(context_dim)
self.global_average_pooling = global_average_pooling
if num_heads_upsample == -1:
num_heads_upsample = num_heads
if num_heads == -1:
assert num_head_channels != -1, 'Either num_heads or num_head_channels has to be set'
if num_head_channels == -1:
assert num_heads != -1, 'Either num_heads or num_head_channels has to be set'
self.dims = dims
self.image_size = image_size
self.in_channels = in_channels
self.model_channels = model_channels
if isinstance(transformer_depth, int):
transformer_depth = len(channel_mult) * [transformer_depth]
if transformer_depth_middle is None:
transformer_depth_middle = transformer_depth[-1]
if isinstance(num_res_blocks, int):
self.num_res_blocks = len(channel_mult) * [num_res_blocks]
else:
if len(num_res_blocks) != len(channel_mult):
raise ValueError("provide num_res_blocks either as an int (globally constant) or "
"as a list/tuple (per-level) with the same length as channel_mult")
self.num_res_blocks = num_res_blocks
if disable_self_attentions is not None:
# should be a list of booleans, indicating whether to disable self-attention in TransformerBlocks or not
assert len(disable_self_attentions) == len(channel_mult)
if num_attention_blocks is not None:
assert len(num_attention_blocks) == len(self.num_res_blocks)
assert all(map(lambda i: self.num_res_blocks[i] >= num_attention_blocks[i], range(
len(num_attention_blocks))))
print(f"Constructor of UNetModel received num_attention_blocks={num_attention_blocks}. "
f"This option has LESS priority than attention_resolutions {attention_resolutions}, "
f"i.e., in cases where num_attention_blocks[i] > 0 but 2**i not in attention_resolutions, "
f"attention will still not be set.")
self.attention_resolutions = attention_resolutions
self.dropout = dropout
self.channel_mult = channel_mult
self.conv_resample = conv_resample
self.num_classes = num_classes
self.use_checkpoint = use_checkpoint
self.dtype = th.float16 if use_fp16 else th.float32
self.dtype = torch.float16 if use_fp16 else torch.float32
self.num_heads = num_heads
self.num_head_channels = num_head_channels
self.num_heads_upsample = num_heads_upsample
@ -188,36 +126,54 @@ class ControlNet(nn.Module):
time_embed_dim = model_channels * 4
self.time_embed = nn.Sequential(
linear(model_channels, time_embed_dim),
linear(model_channels, time_embed_dim, dtype=self.dtype, device=device),
nn.SiLU(),
linear(time_embed_dim, time_embed_dim),
linear(time_embed_dim, time_embed_dim, dtype=self.dtype, device=device),
)
if self.num_classes is not None:
if isinstance(self.num_classes, int):
self.label_emb = nn.Embedding(num_classes, time_embed_dim)
elif self.num_classes == "continuous":
print("setting up linear c_adm embedding layer")
self.label_emb = nn.Linear(1, time_embed_dim)
elif self.num_classes == "sequential":
assert adm_in_channels is not None
self.label_emb = nn.Sequential(
nn.Sequential(
linear(adm_in_channels, time_embed_dim, dtype=self.dtype, device=device),
nn.SiLU(),
linear(time_embed_dim, time_embed_dim, dtype=self.dtype, device=device),
)
)
else:
raise ValueError()
self.input_blocks = nn.ModuleList(
[
TimestepEmbedSequential(
conv_nd(dims, in_channels, model_channels, 3, padding=1)
conv_nd(dims, in_channels, model_channels, 3, padding=1, dtype=self.dtype, device=device)
)
]
)
self.zero_convs = nn.ModuleList([self.make_zero_conv(model_channels)])
self.input_hint_block = TimestepEmbedSequential(
conv_nd(dims, hint_channels, 16, 3, padding=1),
nn.SiLU(),
conv_nd(dims, 16, 16, 3, padding=1),
nn.SiLU(),
conv_nd(dims, 16, 32, 3, padding=1, stride=2),
nn.SiLU(),
conv_nd(dims, 32, 32, 3, padding=1),
nn.SiLU(),
conv_nd(dims, 32, 96, 3, padding=1, stride=2),
nn.SiLU(),
conv_nd(dims, 96, 96, 3, padding=1),
nn.SiLU(),
conv_nd(dims, 96, 256, 3, padding=1, stride=2),
nn.SiLU(),
zero_module(conv_nd(dims, 256, model_channels, 3, padding=1))
conv_nd(dims, hint_channels, 16, 3, padding=1),
nn.SiLU(),
conv_nd(dims, 16, 16, 3, padding=1),
nn.SiLU(),
conv_nd(dims, 16, 32, 3, padding=1, stride=2),
nn.SiLU(),
conv_nd(dims, 32, 32, 3, padding=1),
nn.SiLU(),
conv_nd(dims, 32, 96, 3, padding=1, stride=2),
nn.SiLU(),
conv_nd(dims, 96, 96, 3, padding=1),
nn.SiLU(),
conv_nd(dims, 96, 256, 3, padding=1, stride=2),
nn.SiLU(),
zero_module(conv_nd(dims, 256, model_channels, 3, padding=1))
)
self._feature_size = model_channels
@ -234,7 +190,7 @@ class ControlNet(nn.Module):
out_channels=mult * model_channels,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
use_scale_shift_norm=use_scale_shift_norm
)
]
ch = mult * model_channels
@ -254,14 +210,8 @@ class ControlNet(nn.Module):
if not exists(num_attention_blocks) or nr < num_attention_blocks[level]:
layers.append(
AttentionBlock(
ch,
use_checkpoint=use_checkpoint,
num_heads=num_heads,
num_head_channels=dim_head,
use_new_attention_order=use_new_attention_order,
) if not use_spatial_transformer else SpatialTransformer(
ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim,
SpatialTransformer(
ch, num_heads, dim_head, depth=transformer_depth[level], context_dim=context_dim,
disable_self_attn=disabled_sa, use_linear=use_linear_in_transformer,
use_checkpoint=use_checkpoint
)
@ -282,7 +232,7 @@ class ControlNet(nn.Module):
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
down=True,
down=True
)
if resblock_updown
else Downsample(
@ -311,27 +261,20 @@ class ControlNet(nn.Module):
dropout,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
),
AttentionBlock(
ch,
use_checkpoint=use_checkpoint,
num_heads=num_heads,
num_head_channels=dim_head,
use_new_attention_order=use_new_attention_order,
# always uses a self-attn
) if not use_spatial_transformer else SpatialTransformer(
ch, num_heads, dim_head, depth=transformer_depth, context_dim=context_dim,
disable_self_attn=disable_middle_self_attn, use_linear=use_linear_in_transformer,
use_checkpoint=use_checkpoint
use_scale_shift_norm=use_scale_shift_norm
),
SpatialTransformer( # always uses a self-attn
ch, num_heads, dim_head, depth=transformer_depth_middle, context_dim=context_dim,
disable_self_attn=disable_middle_self_attn, use_linear=use_linear_in_transformer,
use_checkpoint=use_checkpoint
),
ResBlock(
ch,
time_embed_dim,
dropout,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
use_scale_shift_norm=use_scale_shift_norm
),
)
self.middle_block_out = self.make_zero_conv(ch)
@ -339,24 +282,29 @@ class ControlNet(nn.Module):
def make_zero_conv(self, channels):
return TimestepEmbedSequential(zero_module(conv_nd(self.dims, channels, channels, 1, padding=0)))
def align(self, hint, h, w):
b, c, h1, w1 = hint.shape
if h != h1 or w != w1:
return align(hint, (h, w))
return hint
def forward(self, x, hint, timesteps, context, **kwargs):
t_emb = cond_cast_unet(timestep_embedding(timesteps, self.model_channels, repeat_only=False))
def forward(self, x, hint, timesteps, context, y=None, **kwargs):
original_type = x.dtype
x = x.to(self.dtype)
hint = hint.to(self.dtype)
timesteps = timesteps.to(self.dtype)
context = context.to(self.dtype)
if y is not None:
y = y.to(self.dtype)
t_emb = timestep_embedding(timesteps, self.model_channels, repeat_only=False).to(self.dtype)
emb = self.time_embed(t_emb)
guided_hint = self.input_hint_block(cond_cast_unet(hint), emb, context)
outs = []
h1, w1 = x.shape[-2:]
guided_hint = self.align(guided_hint, h1, w1)
h = x.type(self.dtype)
guided_hint = self.input_hint_block(hint, emb, context)
outs = []
if self.num_classes is not None:
assert y.shape[0] == x.shape[0]
emb = emb + self.label_emb(y)
h = x
for module, zero_conv in zip(self.input_blocks, self.zero_convs):
if guided_hint is not None:
h = module(h, emb, context)
@ -369,4 +317,6 @@ class ControlNet(nn.Module):
h = self.middle_block(h, emb, context)
outs.append(self.middle_block_out(h, emb, context))
outs = [o.to(original_type) for o in outs]
return outs

393
scripts/controlmodel_ipadapter.py Normal file
View File

@ -0,0 +1,393 @@
import math
import torch
import torch.nn as nn
# attention_channels of input, output, middle
SD_V12_CHANNELS = [320] * 4 + [640] * 4 + [1280] * 4 + [1280] * 6 + [640] * 6 + [320] * 6 + [1280] * 2
SD_XL_CHANNELS = [640] * 8 + [1280] * 40 + [1280] * 60 + [640] * 12 + [1280] * 20
class ImageProjModel(torch.nn.Module):
"""Projection Model"""
def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024, clip_extra_context_tokens=4):
super().__init__()
self.cross_attention_dim = cross_attention_dim
self.clip_extra_context_tokens = clip_extra_context_tokens
self.proj = torch.nn.Linear(clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim)
self.norm = torch.nn.LayerNorm(cross_attention_dim)
def forward(self, image_embeds):
embeds = image_embeds
clip_extra_context_tokens = self.proj(embeds).reshape(-1, self.clip_extra_context_tokens,
self.cross_attention_dim)
clip_extra_context_tokens = self.norm(clip_extra_context_tokens)
return clip_extra_context_tokens
# Cross Attention to_k, to_v for IPAdapter
class To_KV(torch.nn.Module):
def __init__(self, cross_attention_dim):
super().__init__()
channels = SD_XL_CHANNELS if cross_attention_dim == 2048 else SD_V12_CHANNELS
self.to_kvs = torch.nn.ModuleList(
[torch.nn.Linear(cross_attention_dim, channel, bias=False) for channel in channels])
def load_state_dict(self, state_dict):
# input -> output -> middle
for i, key in enumerate(state_dict.keys()):
self.to_kvs[i].weight.data = state_dict[key]
def FeedForward(dim, mult=4):
inner_dim = int(dim * mult)
return nn.Sequential(
nn.LayerNorm(dim),
nn.Linear(dim, inner_dim, bias=False),
nn.GELU(),
nn.Linear(inner_dim, dim, bias=False),
)
def reshape_tensor(x, heads):
bs, length, width = x.shape
#(bs, length, width) --> (bs, length, n_heads, dim_per_head)
x = x.view(bs, length, heads, -1)
# (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
x = x.transpose(1, 2)
# (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
x = x.reshape(bs, heads, length, -1)
return x
class PerceiverAttention(nn.Module):
def __init__(self, *, dim, dim_head=64, heads=8):
super().__init__()
self.scale = dim_head**-0.5
self.dim_head = dim_head
self.heads = heads
inner_dim = dim_head * heads
self.norm1 = nn.LayerNorm(dim)
self.norm2 = nn.LayerNorm(dim)
self.to_q = nn.Linear(dim, inner_dim, bias=False)
self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
self.to_out = nn.Linear(inner_dim, dim, bias=False)
def forward(self, x, latents):
"""
Args:
x (torch.Tensor): image features
shape (b, n1, D)
latent (torch.Tensor): latent features
shape (b, n2, D)
"""
x = self.norm1(x)
latents = self.norm2(latents)
b, l, _ = latents.shape
q = self.to_q(latents)
kv_input = torch.cat((x, latents), dim=-2)
k, v = self.to_kv(kv_input).chunk(2, dim=-1)
q = reshape_tensor(q, self.heads)
k = reshape_tensor(k, self.heads)
v = reshape_tensor(v, self.heads)
# attention
scale = 1 / math.sqrt(math.sqrt(self.dim_head))
weight = (q * scale) @ (k * scale).transpose(-2, -1) # More stable with f16 than dividing afterwards
weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
out = weight @ v
out = out.permute(0, 2, 1, 3).reshape(b, l, -1)
return self.to_out(out)
class Resampler(nn.Module):
def __init__(
self,
dim=1024,
depth=8,
dim_head=64,
heads=16,
num_queries=8,
embedding_dim=768,
output_dim=1024,
ff_mult=4,
):
super().__init__()
self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim**0.5)
self.proj_in = nn.Linear(embedding_dim, dim)
self.proj_out = nn.Linear(dim, output_dim)
self.norm_out = nn.LayerNorm(output_dim)
self.layers = nn.ModuleList([])
for _ in range(depth):
self.layers.append(
nn.ModuleList(
[
PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
FeedForward(dim=dim, mult=ff_mult),
]
)
)
def forward(self, x):
latents = self.latents.repeat(x.size(0), 1, 1)
x = self.proj_in(x)
for attn, ff in self.layers:
latents = attn(x, latents) + latents
latents = ff(latents) + latents
latents = self.proj_out(latents)
return self.norm_out(latents)
class IPAdapterModel(torch.nn.Module):
def __init__(self, state_dict, clip_embeddings_dim, is_plus):
super().__init__()
self.device = "cpu"
# cross_attention_dim is equal to text_encoder output
self.cross_attention_dim = state_dict["ip_adapter"]["1.to_k_ip.weight"].shape[1]
self.is_plus = is_plus
if self.is_plus:
self.clip_extra_context_tokens = 16
self.image_proj_model = Resampler(
dim=self.cross_attention_dim,
depth=4,
dim_head=64,
heads=12,
num_queries=self.clip_extra_context_tokens,
embedding_dim=clip_embeddings_dim,
output_dim=self.cross_attention_dim,
ff_mult=4
)
else:
self.clip_extra_context_tokens = state_dict["image_proj"]["proj.weight"].shape[0] // self.cross_attention_dim
self.image_proj_model = ImageProjModel(
cross_attention_dim=self.cross_attention_dim,
clip_embeddings_dim=clip_embeddings_dim,
clip_extra_context_tokens=self.clip_extra_context_tokens
)
self.load_ip_adapter(state_dict)
def load_ip_adapter(self, state_dict):
self.image_proj_model.load_state_dict(state_dict["image_proj"])
self.ip_layers = To_KV(self.cross_attention_dim)
self.ip_layers.load_state_dict(state_dict["ip_adapter"])
@torch.inference_mode()
def get_image_embeds(self, clip_vision_output):
self.image_proj_model.cpu()
if self.is_plus:
cond = self.image_proj_model(clip_vision_output['hidden_states'][-2].to(device='cpu', dtype=torch.float32))
uncond = self.image_proj_model(torch.zeros_like(clip_vision_output['hidden_states'][-2].to(device='cpu', dtype=torch.float32)))
return cond, uncond
clip_image_embeds = clip_vision_output['image_embeds'].to(device='cpu', dtype=torch.float32)
image_prompt_embeds = self.image_proj_model(clip_image_embeds)
# input zero vector for unconditional.
uncond_image_prompt_embeds = self.image_proj_model(torch.zeros_like(clip_image_embeds))
return image_prompt_embeds, uncond_image_prompt_embeds
def get_block(model, flag):
return {
'input': model.input_blocks, 'middle': [model.middle_block], 'output': model.output_blocks
}[flag]
def attn_forward_hacked(self, x, context=None, **kwargs):
batch_size, sequence_length, inner_dim = x.shape
h = self.heads
head_dim = inner_dim // h
if context is None:
context = x
q = self.to_q(x)
k = self.to_k(context)
v = self.to_v(context)
del context
q, k, v = map(
lambda t: t.view(batch_size, -1, h, head_dim).transpose(1, 2),
(q, k, v),
)
out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=None, dropout_p=0.0, is_causal=False)
out = out.transpose(1, 2).reshape(batch_size, -1, h * head_dim)
del k, v
for f in self.ipadapter_hacks:
out = out + f(self, x, q)
del q, x
return self.to_out(out)
all_hacks = {}
current_model = None
def hack_blk(block, function, type):
if not hasattr(block, 'ipadapter_hacks'):
block.ipadapter_hacks = []
if len(block.ipadapter_hacks) == 0:
all_hacks[block] = block.forward
block.forward = attn_forward_hacked.__get__(block, type)
block.ipadapter_hacks.append(function)
return
def set_model_attn2_replace(model, function, flag, id):
from ldm.modules.attention import CrossAttention
block = get_block(model, flag)[id][1].transformer_blocks[0].attn2
hack_blk(block, function, CrossAttention)
return
def set_model_patch_replace(model, function, flag, id, trans_id):
from sgm.modules.attention import CrossAttention
blk = get_block(model, flag)
block = blk[id][1].transformer_blocks[trans_id].attn2
hack_blk(block, function, CrossAttention)
return
def clear_all_ip_adapter():
global all_hacks, current_model
for k, v in all_hacks.items():
k.forward = v
k.ipadapter_hacks = []
all_hacks = {}
current_model = None
return
class PlugableIPAdapter(torch.nn.Module):
def __init__(self, state_dict, clip_embeddings_dim, is_plus):
super().__init__()
self.sdxl = clip_embeddings_dim == 1280 and not is_plus
self.is_plus = is_plus
self.ipadapter = IPAdapterModel(state_dict, clip_embeddings_dim=clip_embeddings_dim, is_plus=is_plus)
self.disable_memory_management = True
self.dtype = None
self.weight = 1.0
self.cache = {}
self.p_start = 0.0
self.p_end = 1.0
return
def reset(self):
self.cache = {}
return
@torch.no_grad()
def hook(self, model, clip_vision_output, weight, start, end, dtype=torch.float32):
global current_model
current_model = model
self.p_start = start
self.p_end = end
self.cache = {}
self.weight = weight
device = torch.device('cpu')
self.dtype = dtype
self.ipadapter.to(device, dtype=self.dtype)
self.image_emb, self.uncond_image_emb = self.ipadapter.get_image_embeds(clip_vision_output)
self.image_emb = self.image_emb.to(device, dtype=self.dtype)
self.uncond_image_emb = self.uncond_image_emb.to(device, dtype=self.dtype)
# From https://github.com/laksjdjf/IPAdapter-ComfyUI
if not self.sdxl:
number = 0 # index of to_kvs
for id in [1, 2, 4, 5, 7, 8]: # id of input_blocks that have cross attention
set_model_attn2_replace(model, self.patch_forward(number), "input", id)
number += 1
for id in [3, 4, 5, 6, 7, 8, 9, 10, 11]: # id of output_blocks that have cross attention
set_model_attn2_replace(model, self.patch_forward(number), "output", id)
number += 1
set_model_attn2_replace(model, self.patch_forward(number), "middle", 0)
else:
number = 0
for id in [4, 5, 7, 8]: # id of input_blocks that have cross attention
block_indices = range(2) if id in [4, 5] else range(10) # transformer_depth
for index in block_indices:
set_model_patch_replace(model, self.patch_forward(number), "input", id, index)
number += 1
for id in range(6): # id of output_blocks that have cross attention
block_indices = range(2) if id in [3, 4, 5] else range(10) # transformer_depth
for index in block_indices:
set_model_patch_replace(model, self.patch_forward(number), "output", id, index)
number += 1
for index in range(10):
set_model_patch_replace(model, self.patch_forward(number), "middle", 0, index)
number += 1
return
def call_ip(self, number, feat, device):
if number in self.cache:
return self.cache[number]
else:
ip = self.ipadapter.ip_layers.to_kvs[number](feat).to(device)
self.cache[number] = ip
return ip
@torch.no_grad()
def patch_forward(self, number):
@torch.no_grad()
def forward(attn_blk, x, q):
batch_size, sequence_length, inner_dim = x.shape
h = attn_blk.heads
head_dim = inner_dim // h
current_sampling_percent = getattr(current_model, 'current_sampling_percent', 0.5)
if current_sampling_percent < self.p_start or current_sampling_percent > self.p_end:
return 0
cond_mark = current_model.cond_mark[:, :, :, 0].to(self.image_emb)
cond_uncond_image_emb = self.image_emb * cond_mark + self.uncond_image_emb * (1 - cond_mark)
ip_k = self.call_ip(number * 2, cond_uncond_image_emb, device=q.device)
ip_v = self.call_ip(number * 2 + 1, cond_uncond_image_emb, device=q.device)
ip_k, ip_v = map(
lambda t: t.view(batch_size, -1, h, head_dim).transpose(1, 2),
(ip_k, ip_v),
)
ip_out = torch.nn.functional.scaled_dot_product_attention(q, ip_k, ip_v, attn_mask=None, dropout_p=0.0, is_causal=False)
ip_out = ip_out.transpose(1, 2).reshape(batch_size, -1, h * head_dim)
return ip_out * self.weight
return forward

337
scripts/controlnet.py
View File

@ -12,12 +12,13 @@ import gradio as gr
from einops import rearrange
from scripts import global_state, hook, external_code, processor, batch_hijack, controlnet_version, utils
from scripts.controlnet_ui import controlnet_ui_group
from scripts.cldm import PlugableControlModel
from scripts.controlnet_lora import bind_control_lora, unbind_control_lora
from scripts.processor import *
from scripts.adapter import PlugableAdapter
from scripts.adapter import Adapter, StyleAdapter, Adapter_light
from scripts.controlnet_lllite import PlugableControlLLLite, clear_all_lllite
from scripts.controlmodel_ipadapter import PlugableIPAdapter, clear_all_ip_adapter
from scripts.utils import load_state_dict, get_unique_axis0
from scripts.hook import ControlParams, UnetHook, ControlModelType
from scripts.hook import ControlParams, UnetHook, ControlModelType, HackedImageRNG
from scripts.controlnet_ui.controlnet_ui_group import ControlNetUiGroup, UiControlNetUnit
from scripts.logging import logger
from modules.processing import StableDiffusionProcessingImg2Img, StableDiffusionProcessingTxt2Img
@ -32,6 +33,8 @@ from pathlib import Path
from PIL import Image, ImageFilter, ImageOps
from scripts.lvminthin import lvmin_thin, nake_nms
from scripts.processor import model_free_preprocessors
from scripts.controlnet_model_guess import build_model_by_guess
gradio_compat = True
try:
@ -49,6 +52,11 @@ gradio_tempfile_path = os.path.join(tempfile.gettempdir(), 'gradio')
os.makedirs(gradio_tempfile_path, exist_ok=True)
def clear_all_secondary_control_models():
clear_all_lllite()
clear_all_ip_adapter()
def find_closest_lora_model_name(search: str):
if not search:
return None
@ -228,6 +236,7 @@ class Script(scripts.Script, metaclass=(
self.enabled_units = []
self.detected_map = []
self.post_processors = []
self.noise_modifier = None
batch_hijack.instance.process_batch_callbacks.append(self.batch_tab_process)
batch_hijack.instance.process_batch_each_callbacks.append(self.batch_tab_process_each)
batch_hijack.instance.postprocess_batch_each_callbacks.insert(0, self.batch_tab_postprocess_each)
@ -266,7 +275,7 @@ class Script(scripts.Script, metaclass=(
infotext = Infotext()
controls = ()
max_models = shared.opts.data.get("control_net_max_models_num", 1)
max_models = shared.opts.data.get("control_net_unit_count", 3)
elem_id_tabname = ("img2img" if is_img2img else "txt2img") + "_controlnet"
with gr.Group(elem_id=elem_id_tabname):
with gr.Accordion(f"ControlNet {controlnet_version.version_flag}", open = False, elem_id="controlnet"):
@ -297,7 +306,7 @@ class Script(scripts.Script, metaclass=(
devices.torch_gc()
@staticmethod
def load_control_model(p, unet, model, lowvram):
def load_control_model(p, unet, model):
if model in Script.model_cache:
logger.info(f"Loading model from cache: {model}")
return Script.model_cache[model]
@ -308,7 +317,7 @@ class Script(scripts.Script, metaclass=(
gc.collect()
devices.torch_gc()
model_net = Script.build_control_model(p, unet, model, lowvram)
model_net = Script.build_control_model(p, unet, model)
if shared.opts.data.get("control_net_model_cache_size", 2) > 0:
Script.model_cache[model] = model_net
@ -316,7 +325,7 @@ class Script(scripts.Script, metaclass=(
return model_net
@staticmethod
def build_control_model(p, unet, model, lowvram):
def build_control_model(p, unet, model):
if model is None or model == 'None':
raise RuntimeError(f"You have not selected any ControlNet Model.")
@ -337,67 +346,8 @@ class Script(scripts.Script, metaclass=(
logger.info(f"Loading model: {model}")
state_dict = load_state_dict(model_path)
network_module = PlugableControlModel
network_config = shared.opts.data.get("control_net_model_config", global_state.default_conf)
if not os.path.isabs(network_config):
network_config = os.path.join(global_state.script_dir, network_config)
if any([k.startswith("body.") or k == 'style_embedding' for k, v in state_dict.items()]):
# adapter model
network_module = PlugableAdapter
network_config = shared.opts.data.get("control_net_model_adapter_config", global_state.default_conf_adapter)
if not os.path.isabs(network_config):
network_config = os.path.join(global_state.script_dir, network_config)
model_path = os.path.abspath(model_path)
model_stem = Path(model_path).stem
model_dir_name = os.path.dirname(model_path)
possible_config_filenames = [
os.path.join(model_dir_name, model_stem + ".yaml"),
os.path.join(global_state.script_dir, 'models', model_stem + ".yaml"),
os.path.join(model_dir_name, model_stem.replace('_fp16', '') + ".yaml"),
os.path.join(global_state.script_dir, 'models', model_stem.replace('_fp16', '') + ".yaml"),
os.path.join(model_dir_name, model_stem.replace('_diff', '') + ".yaml"),
os.path.join(global_state.script_dir, 'models', model_stem.replace('_diff', '') + ".yaml"),
os.path.join(model_dir_name, model_stem.replace('-fp16', '') + ".yaml"),
os.path.join(global_state.script_dir, 'models', model_stem.replace('-fp16', '') + ".yaml"),
os.path.join(model_dir_name, model_stem.replace('-diff', '') + ".yaml"),
os.path.join(global_state.script_dir, 'models', model_stem.replace('-diff', '') + ".yaml")
]
override_config = possible_config_filenames[0]
for possible_config_filename in possible_config_filenames:
if os.path.exists(possible_config_filename):
override_config = possible_config_filename
break
if 'v11' in model_stem.lower() or 'shuffle' in model_stem.lower():
assert os.path.exists(override_config), f'Error: The model config {override_config} is missing. ControlNet 1.1 must have configs.'
if os.path.exists(override_config):
network_config = override_config
else:
# Note: This error is triggered in unittest, but not caught.
# TODO: Replace `print` with `logger.error`.
print(f'ERROR: ControlNet cannot find model config [{override_config}] \n'
f'ERROR: ControlNet will use a WRONG config [{network_config}] to load your model. \n'
f'ERROR: The WRONG config may not match your model. The generated results can be bad. \n'
f'ERROR: You are using a ControlNet model [{model_stem}] without correct YAML config file. \n'
f'ERROR: The performance of this model may be worse than your expectation. \n'
f'ERROR: If this model cannot get good results, the reason is that you do not have a YAML file for the model. \n'
f'Solution: Please download YAML file, or ask your model provider to provide [{override_config}] for you to download.\n'
f'Hint: You can take a look at [{os.path.join(global_state.script_dir, "models")}] to find many existing YAML files.\n')
logger.info(f"Loading config: {network_config}")
network = network_module(
state_dict=state_dict,
config_path=network_config,
lowvram=lowvram,
base_model=unet,
)
network.to(p.sd_model.device, dtype=p.sd_model.dtype)
network = build_model_by_guess(state_dict, unet, model_path)
network.to('cpu', dtype=p.sd_model.dtype)
logger.info(f"ControlNet model {model} loaded.")
return network
@ -482,41 +432,44 @@ class Script(scripts.Script, metaclass=(
inpaint_mask = x[:, :, 3]
x = x[:, :, 0:3]
new_size_is_smaller = (size[0] * size[1]) < (x.shape[0] * x.shape[1])
new_size_is_bigger = (size[0] * size[1]) > (x.shape[0] * x.shape[1])
unique_color_count = len(get_unique_axis0(x.reshape(-1, x.shape[2])))
is_one_pixel_edge = False
is_binary = False
if unique_color_count == 2:
is_binary = np.min(x) < 16 and np.max(x) > 240
if is_binary:
xc = x
xc = cv2.erode(xc, np.ones(shape=(3, 3), dtype=np.uint8), iterations=1)
xc = cv2.dilate(xc, np.ones(shape=(3, 3), dtype=np.uint8), iterations=1)
one_pixel_edge_count = np.where(xc < x)[0].shape[0]
all_edge_count = np.where(x > 127)[0].shape[0]
is_one_pixel_edge = one_pixel_edge_count * 2 > all_edge_count
if x.shape[0] != size[1] or x.shape[1] != size[0]:
new_size_is_smaller = (size[0] * size[1]) < (x.shape[0] * x.shape[1])
new_size_is_bigger = (size[0] * size[1]) > (x.shape[0] * x.shape[1])
unique_color_count = len(get_unique_axis0(x.reshape(-1, x.shape[2])))
is_one_pixel_edge = False
is_binary = False
if unique_color_count == 2:
is_binary = np.min(x) < 16 and np.max(x) > 240
if is_binary:
xc = x
xc = cv2.erode(xc, np.ones(shape=(3, 3), dtype=np.uint8), iterations=1)
xc = cv2.dilate(xc, np.ones(shape=(3, 3), dtype=np.uint8), iterations=1)
one_pixel_edge_count = np.where(xc < x)[0].shape[0]
all_edge_count = np.where(x > 127)[0].shape[0]
is_one_pixel_edge = one_pixel_edge_count * 2 > all_edge_count
if 2 < unique_color_count < 200:
interpolation = cv2.INTER_NEAREST
elif new_size_is_smaller:
interpolation = cv2.INTER_AREA
else:
interpolation = cv2.INTER_CUBIC # Must be CUBIC because we now use nms. NEVER CHANGE THIS
y = cv2.resize(x, size, interpolation=interpolation)
if inpaint_mask is not None:
inpaint_mask = cv2.resize(inpaint_mask, size, interpolation=interpolation)
if is_binary:
y = np.mean(y.astype(np.float32), axis=2).clip(0, 255).astype(np.uint8)
if is_one_pixel_edge:
y = nake_nms(y)
_, y = cv2.threshold(y, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
y = lvmin_thin(y, prunings=new_size_is_bigger)
if 2 < unique_color_count < 200:
interpolation = cv2.INTER_NEAREST
elif new_size_is_smaller:
interpolation = cv2.INTER_AREA
else:
_, y = cv2.threshold(y, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
y = np.stack([y] * 3, axis=2)
interpolation = cv2.INTER_CUBIC # Must be CUBIC because we now use nms. NEVER CHANGE THIS
y = cv2.resize(x, size, interpolation=interpolation)
if inpaint_mask is not None:
inpaint_mask = cv2.resize(inpaint_mask, size, interpolation=interpolation)
if is_binary:
y = np.mean(y.astype(np.float32), axis=2).clip(0, 255).astype(np.uint8)
if is_one_pixel_edge:
y = nake_nms(y)
_, y = cv2.threshold(y, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
y = lvmin_thin(y, prunings=new_size_is_bigger)
else:
_, y = cv2.threshold(y, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)
y = np.stack([y] * 3, axis=2)
else:
y = x
if inpaint_mask is not None:
inpaint_mask = (inpaint_mask > 127).astype(np.float32) * 255.0
@ -691,12 +644,16 @@ class Script(scripts.Script, metaclass=(
"""
sd_ldm = p.sd_model
unet = sd_ldm.model.diffusion_model
self.noise_modifier = None
setattr(p, 'controlnet_initial_noise_modifier', None)
setattr(p, 'controlnet_control_loras', [])
if self.latest_network is not None:
# always restore (~0.05s)
self.latest_network.restore(unet)
self.latest_network.restore()
# always clear (~0.05s)
clear_all_secondary_control_models()
if not batch_hijack.instance.is_batch:
self.enabled_units = Script.get_enabled_units(p)
@ -704,11 +661,6 @@ class Script(scripts.Script, metaclass=(
if len(self.enabled_units) == 0:
self.latest_network = None
return
is_sdxl = getattr(p.sd_model, 'is_sdxl', False)
if is_sdxl:
logger.warning('ControlNet does not support SDXL -- disabling')
return
detected_maps = []
forward_params = []
@ -718,6 +670,9 @@ class Script(scripts.Script, metaclass=(
if self.latest_model_hash != p.sd_model.sd_model_hash:
Script.clear_control_model_cache()
for idx, unit in enumerate(self.enabled_units):
unit.module = global_state.get_module_basename(unit.module)
# unload unused preproc
module_list = [unit.module for unit in self.enabled_units]
for key in self.unloadable:
@ -728,16 +683,20 @@ class Script(scripts.Script, metaclass=(
for idx, unit in enumerate(self.enabled_units):
Script.bound_check_params(unit)
unit.module = global_state.get_module_basename(unit.module)
resize_mode = external_code.resize_mode_from_value(unit.resize_mode)
control_mode = external_code.control_mode_from_value(unit.control_mode)
if unit.module in model_free_preprocessors:
model_net = None
else:
model_net = Script.load_control_model(p, unet, unit.model, unit.low_vram)
model_net = Script.load_control_model(p, unet, unit.model)
model_net.reset()
if getattr(model_net, 'is_control_lora', False):
control_lora = model_net.control_model
bind_control_lora(unet, control_lora)
p.controlnet_control_loras.append(control_lora)
input_image, image_from_a1111 = Script.choose_input_image(p, unit, idx)
if image_from_a1111:
a1111_i2i_resize_mode = getattr(p, "resize_mode", None)
@ -814,6 +773,25 @@ class Script(scripts.Script, metaclass=(
# inpaint_only+lama is special and required outpaint fix
_, input_image = Script.detectmap_proc(input_image, unit.module, resize_mode, hr_y, hr_x)
control_model_type = ControlModelType.ControlNet
global_average_pooling = False
if 'reference' in unit.module:
control_model_type = ControlModelType.AttentionInjection
elif 'revision' in unit.module:
control_model_type = ControlModelType.ReVision
elif hasattr(model_net, 'control_model') and (isinstance(model_net.control_model, Adapter) or isinstance(model_net.control_model, Adapter_light)):
control_model_type = ControlModelType.T2I_Adapter
elif hasattr(model_net, 'control_model') and isinstance(model_net.control_model, StyleAdapter):
control_model_type = ControlModelType.T2I_StyleAdapter
elif isinstance(model_net, PlugableIPAdapter):
control_model_type = ControlModelType.IPAdapter
elif isinstance(model_net, PlugableControlLLLite):
control_model_type = ControlModelType.Controlllite
if control_model_type is ControlModelType.ControlNet:
global_average_pooling = model_net.control_model.global_average_pooling
preprocessor_resolution = unit.processor_res
if unit.pixel_perfect:
preprocessor_resolution = external_code.pixel_perfect_resolution(
@ -838,12 +816,6 @@ class Script(scripts.Script, metaclass=(
thr_b=unit.threshold_b,
)
if unit.module == "none" and "style" in unit.model:
detected_map_bytes = detected_map[:,:,0].tobytes()
detected_map = np.ndarray((round(input_image.shape[0]/4),input_image.shape[1]),dtype="float32",buffer=detected_map_bytes)
detected_map = torch.Tensor(detected_map).to(devices.get_device_for("controlnet"))
is_image = False
if high_res_fix:
if is_image:
hr_control, hr_detected_map = Script.detectmap_proc(detected_map, unit.module, resize_mode, hr_y, hr_x)
@ -858,25 +830,13 @@ class Script(scripts.Script, metaclass=(
detected_maps.append((detected_map, unit.module))
else:
control = detected_map
if unit.module == 'clip_vision':
detected_maps.append((processor.clip_vision_visualization(detected_map), unit.module))
detected_maps.append((input_image, unit.module))
control_model_type = ControlModelType.ControlNet
if control_model_type == ControlModelType.T2I_StyleAdapter:
control = control['last_hidden_state']
if isinstance(model_net, PlugableAdapter):
control_model_type = ControlModelType.T2I_Adapter
if getattr(model_net, "target", None) == "scripts.adapter.StyleAdapter":
control_model_type = ControlModelType.T2I_StyleAdapter
if 'reference' in unit.module:
control_model_type = ControlModelType.AttentionInjection
global_average_pooling = False
if model_net is not None:
if model_net.config.model.params.get("global_average_pooling", False):
global_average_pooling = True
if control_model_type == ControlModelType.ReVision:
control = control['image_embeds']
preprocessor_dict = dict(
name=unit.module,
@ -928,16 +888,100 @@ class Script(scripts.Script, metaclass=(
post_processors.append(inpaint_only_post_processing)
if 'recolor' in unit.module:
final_feed = hr_control if hr_control is not None else control
final_feed = final_feed.detach().cpu().numpy()
final_feed = np.ascontiguousarray(final_feed).copy()
final_feed = final_feed[0, 0, :, :].astype(np.float32)
final_feed = (final_feed * 255).clip(0, 255).astype(np.uint8)
Hfeed, Wfeed = final_feed.shape
if 'luminance' in unit.module:
def recolor_luminance_post_processing(x):
C, H, W = x.shape
if Hfeed != H or Wfeed != W or C != 3:
logger.error('Error: ControlNet find post-processing resolution mismatch. This could be related to other extensions hacked processing.')
return x
h = x.detach().cpu().numpy().transpose((1, 2, 0))
h = (h * 255).clip(0, 255).astype(np.uint8)
h = cv2.cvtColor(h, cv2.COLOR_RGB2LAB)
h[:, :, 0] = final_feed
h = cv2.cvtColor(h, cv2.COLOR_LAB2RGB)
h = (h.astype(np.float32) / 255.0).transpose((2, 0, 1))
y = torch.from_numpy(h).clip(0, 1).to(x)
return y
post_processors.append(recolor_luminance_post_processing)
if 'intensity' in unit.module:
def recolor_intensity_post_processing(x):
C, H, W = x.shape
if Hfeed != H or Wfeed != W or C != 3:
logger.error('Error: ControlNet find post-processing resolution mismatch. This could be related to other extensions hacked processing.')
return x
h = x.detach().cpu().numpy().transpose((1, 2, 0))
h = (h * 255).clip(0, 255).astype(np.uint8)
h = cv2.cvtColor(h, cv2.COLOR_RGB2HSV)
h[:, :, 2] = final_feed
h = cv2.cvtColor(h, cv2.COLOR_HSV2RGB)
h = (h.astype(np.float32) / 255.0).transpose((2, 0, 1))
y = torch.from_numpy(h).clip(0, 1).to(x)
return y
post_processors.append(recolor_intensity_post_processing)
if '+lama' in unit.module:
forward_param.used_hint_cond_latent = hook.UnetHook.call_vae_using_process(p, control)
setattr(p, 'controlnet_initial_noise_modifier', forward_param.used_hint_cond_latent)
self.noise_modifier = forward_param.used_hint_cond_latent
del model_net
self.latest_network = UnetHook(lowvram=any(unit.low_vram for unit in self.enabled_units))
is_low_vram = any(unit.low_vram for unit in self.enabled_units)
self.latest_network = UnetHook(lowvram=is_low_vram)
self.latest_network.hook(model=unet, sd_ldm=sd_ldm, control_params=forward_params, process=p)
revision_conds = 0
revision_conds_weight = 0
for param in forward_params:
if param.control_model_type == ControlModelType.ReVision:
revision_conds = revision_conds + param.hint_cond * param.weight
revision_conds_weight += param.weight
revision_conds_weight = max(revision_conds_weight, 1e-3)
self.latest_network.global_revision = revision_conds / revision_conds_weight
for param in forward_params:
if param.control_model_type == ControlModelType.IPAdapter:
param.control_model.hook(
model=unet,
clip_vision_output=param.hint_cond,
weight=param.weight,
dtype=torch.float32,
start=param.start_guidance_percent,
end=param.stop_guidance_percent
)
if param.control_model_type == ControlModelType.Controlllite:
param.control_model.hook(
model=unet,
cond=param.hint_cond,
weight=param.weight,
start=param.start_guidance_percent,
end=param.stop_guidance_percent
)
self.detected_map = detected_maps
self.post_processors = post_processors
def before_process_batch(self, p, *args, **kwargs):
if self.noise_modifier is not None:
p.rng = HackedImageRNG(rng=p.rng,
noise_modifier=self.noise_modifier,
sd_model=p.sd_model)
self.noise_modifier = None
return
def postprocess_batch(self, p, *args, **kwargs):
images = kwargs.get('images', [])
for post_processor in self.post_processors:
@ -946,8 +990,15 @@ class Script(scripts.Script, metaclass=(
return
def postprocess(self, p, processed, *args):
clear_all_secondary_control_models()
self.noise_modifier = None
for control_lora in getattr(p, 'controlnet_control_loras', []):
unbind_control_lora(control_lora)
p.controlnet_control_loras = []
self.post_processors = []
setattr(p, 'controlnet_initial_noise_modifier', None)
setattr(p, 'controlnet_vae_cache', None)
processor_params_flag = (', '.join(getattr(processed, 'extra_generation_params', []))).lower()
@ -985,7 +1036,7 @@ class Script(scripts.Script, metaclass=(
])
self.input_image = None
self.latest_network.restore(p.sd_model.model.diffusion_model)
self.latest_network.restore()
self.latest_network = None
self.detected_map.clear()
@ -1017,25 +1068,21 @@ class Script(scripts.Script, metaclass=(
self.input_image = None
if self.latest_network is None: return
self.latest_network.restore(shared.sd_model.model.diffusion_model)
self.latest_network.restore()
self.latest_network = None
self.detected_map.clear()
def on_ui_settings():
section = ('control_net', "ControlNet")
shared.opts.add_option("control_net_model_config", shared.OptionInfo(
global_state.default_conf, "Config file for Control Net models", section=section))
shared.opts.add_option("control_net_model_adapter_config", shared.OptionInfo(
global_state.default_conf_adapter, "Config file for Adapter models", section=section))
shared.opts.add_option("control_net_detectedmap_dir", shared.OptionInfo(
global_state.default_detectedmap_dir, "Directory for detected maps auto saving", section=section))
shared.opts.add_option("control_net_models_path", shared.OptionInfo(
"", "Extra path to scan for ControlNet models (e.g. training output directory)", section=section))
shared.opts.add_option("control_net_modules_path", shared.OptionInfo(
"", "Path to directory containing annotator model directories (requires restart, overrides corresponding command line flag)", section=section))
shared.opts.add_option("control_net_max_models_num", shared.OptionInfo(
3, "Multi ControlNet: Max models amount (requires restart)", gr.Slider, {"minimum": 1, "maximum": 10, "step": 1}, section=section))
shared.opts.add_option("control_net_unit_count", shared.OptionInfo(
3, "Multi-ControlNet: ControlNet unit number (requires restart)", gr.Slider, {"minimum": 1, "maximum": 10, "step": 1}, section=section))
shared.opts.add_option("control_net_model_cache_size", shared.OptionInfo(
1, "Model cache size (requires restart)", gr.Slider, {"minimum": 1, "maximum": 5, "step": 1}, section=section))
shared.opts.add_option("control_net_inpaint_blur_sigma", shared.OptionInfo(

95
scripts/controlnet_diffusers.py Normal file
View File

@ -0,0 +1,95 @@
# https://gist.github.com/takuma104/4adfb3d968d80bea1d18a30c06439242
# 2nd editing by lllyasviel
import torch
# =================#
# UNet Conversion #
# =================#
unet_conversion_map = [
# (stable-diffusion, HF Diffusers)
("time_embed.0.weight", "time_embedding.linear_1.weight"),
("time_embed.0.bias", "time_embedding.linear_1.bias"),
("time_embed.2.weight", "time_embedding.linear_2.weight"),
("time_embed.2.bias", "time_embedding.linear_2.bias"),
("label_emb.0.0.weight", "add_embedding.linear_1.weight"),
("label_emb.0.0.bias", "add_embedding.linear_1.bias"),
("label_emb.0.2.weight", "add_embedding.linear_2.weight"),
("label_emb.0.2.bias", "add_embedding.linear_2.bias"),
("input_blocks.0.0.weight", "conv_in.weight"),
("input_blocks.0.0.bias", "conv_in.bias"),
("middle_block_out.0.weight", "controlnet_mid_block.weight"),
("middle_block_out.0.bias", "controlnet_mid_block.bias"),
]
unet_conversion_map_resnet = [
# (stable-diffusion, HF Diffusers)
("in_layers.0", "norm1"),
("in_layers.2", "conv1"),
("out_layers.0", "norm2"),
("out_layers.3", "conv2"),
("emb_layers.1", "time_emb_proj"),
("skip_connection", "conv_shortcut"),
]
unet_conversion_map_layer = []
# hardcoded number of downblocks and resnets/attentions...
# would need smarter logic for other networks.
for i in range(4):
# loop over downblocks/upblocks
for j in range(10):
# loop over resnets/attentions for downblocks
hf_down_res_prefix = f"down_blocks.{i}.resnets.{j}."
sd_down_res_prefix = f"input_blocks.{3*i + j + 1}.0."
unet_conversion_map_layer.append((sd_down_res_prefix, hf_down_res_prefix))
hf_down_atn_prefix = f"down_blocks.{i}.attentions.{j}."
sd_down_atn_prefix = f"input_blocks.{3 * i + j + 1}.1."
unet_conversion_map_layer.append((sd_down_atn_prefix, hf_down_atn_prefix))
hf_downsample_prefix = f"down_blocks.{i}.downsamplers.0.conv."
sd_downsample_prefix = f"input_blocks.{3 * (i + 1)}.0.op."
unet_conversion_map_layer.append((sd_downsample_prefix, hf_downsample_prefix))
hf_mid_atn_prefix = "mid_block.attentions.0."
sd_mid_atn_prefix = "middle_block.1."
unet_conversion_map_layer.append((sd_mid_atn_prefix, hf_mid_atn_prefix))
for j in range(2):
hf_mid_res_prefix = f"mid_block.resnets.{j}."
sd_mid_res_prefix = f"middle_block.{2*j}."
unet_conversion_map_layer.append((sd_mid_res_prefix, hf_mid_res_prefix))
# controlnet specific
controlnet_cond_embedding_names = ['conv_in'] + [f'blocks.{i}' for i in range(6)] + ['conv_out']
for i, hf_prefix in enumerate(controlnet_cond_embedding_names):
hf_prefix = f"controlnet_cond_embedding.{hf_prefix}."
sd_prefix = f"input_hint_block.{i*2}."
unet_conversion_map_layer.append((sd_prefix, hf_prefix))
for i in range(12):
hf_prefix = f"controlnet_down_blocks.{i}."
sd_prefix = f"zero_convs.{i}.0."
unet_conversion_map_layer.append((sd_prefix, hf_prefix))
def convert_from_diffuser_state_dict(unet_state_dict):
mapping = {k: k for k in unet_state_dict.keys()}
for sd_name, hf_name in unet_conversion_map:
mapping[hf_name] = sd_name
for k, v in mapping.items():
if "resnets" in k:
for sd_part, hf_part in unet_conversion_map_resnet:
v = v.replace(hf_part, sd_part)
mapping[k] = v
for k, v in mapping.items():
for sd_part, hf_part in unet_conversion_map_layer:
v = v.replace(hf_part, sd_part)
mapping[k] = v
new_state_dict = {v: unet_state_dict[k] for k, v in mapping.items() if k in unet_state_dict}
return new_state_dict

218
scripts/controlnet_lllite.py Normal file
View File

@ -0,0 +1,218 @@
# https://github.com/kohya-ss/ControlNet-LLLite-ComfyUI/blob/main/node_control_net_lllite.py
import re
import torch
from modules import devices
class LLLiteModule(torch.nn.Module):
def __init__(
self,
name: str,
is_conv2d: bool,
in_dim: int,
depth: int,
cond_emb_dim: int,
mlp_dim: int,
):
super().__init__()
self.name = name
self.is_conv2d = is_conv2d
self.is_first = False
modules = []
modules.append(torch.nn.Conv2d(3, cond_emb_dim // 2, kernel_size=4, stride=4, padding=0)) # to latent (from VAE) size*2
if depth == 1:
modules.append(torch.nn.ReLU(inplace=True))
modules.append(torch.nn.Conv2d(cond_emb_dim // 2, cond_emb_dim, kernel_size=2, stride=2, padding=0))
elif depth == 2:
modules.append(torch.nn.ReLU(inplace=True))
modules.append(torch.nn.Conv2d(cond_emb_dim // 2, cond_emb_dim, kernel_size=4, stride=4, padding=0))
elif depth == 3:
# kernel size 8は大きすぎるので、4にする / kernel size 8 is too large, so set it to 4
modules.append(torch.nn.ReLU(inplace=True))
modules.append(torch.nn.Conv2d(cond_emb_dim // 2, cond_emb_dim // 2, kernel_size=4, stride=4, padding=0))
modules.append(torch.nn.ReLU(inplace=True))
modules.append(torch.nn.Conv2d(cond_emb_dim // 2, cond_emb_dim, kernel_size=2, stride=2, padding=0))
self.conditioning1 = torch.nn.Sequential(*modules)
if self.is_conv2d:
self.down = torch.nn.Sequential(
torch.nn.Conv2d(in_dim, mlp_dim, kernel_size=1, stride=1, padding=0),
torch.nn.ReLU(inplace=True),
)
self.mid = torch.nn.Sequential(
torch.nn.Conv2d(mlp_dim + cond_emb_dim, mlp_dim, kernel_size=1, stride=1, padding=0),
torch.nn.ReLU(inplace=True),
)
self.up = torch.nn.Sequential(
torch.nn.Conv2d(mlp_dim, in_dim, kernel_size=1, stride=1, padding=0),
)
else:
self.down = torch.nn.Sequential(
torch.nn.Linear(in_dim, mlp_dim),
torch.nn.ReLU(inplace=True),
)
self.mid = torch.nn.Sequential(
torch.nn.Linear(mlp_dim + cond_emb_dim, mlp_dim),
torch.nn.ReLU(inplace=True),
)
self.up = torch.nn.Sequential(
torch.nn.Linear(mlp_dim, in_dim),
)
self.depth = depth
self.cond_image = None
self.cond_emb = None
def set_cond_image(self, cond_image):
self.cond_image = cond_image
self.cond_emb = None
def forward(self, x):
if self.cond_emb is None:
# print(f"cond_emb is None, {self.name}")
cx = self.conditioning1(self.cond_image.to(x.device, dtype=x.dtype))
if not self.is_conv2d:
# reshape / b,c,h,w -> b,h*w,c
n, c, h, w = cx.shape
cx = cx.view(n, c, h * w).permute(0, 2, 1)
self.cond_emb = cx
cx = self.cond_emb
# uncond/condでxはバッチサイズが2倍
if x.shape[0] != cx.shape[0]:
if self.is_conv2d:
cx = cx.repeat(x.shape[0] // cx.shape[0], 1, 1, 1)
else:
# print("x.shape[0] != cx.shape[0]", x.shape[0], cx.shape[0])
cx = cx.repeat(x.shape[0] // cx.shape[0], 1, 1)
try:
cx = torch.cat([cx, self.down(x)], dim=1 if self.is_conv2d else 2)
cx = self.mid(cx)
cx = self.up(cx)
return cx
except RuntimeError as e:
# high-res fix shape mismatch
return 0
all_hack = {}
def clear_all_lllite():
global all_hack
for k, v in all_hack.items():
k.forward = v
k.lllite_list = []
all_hack = {}
return
class PlugableControlLLLite(torch.nn.Module):
def __init__(self, state_dict):
super().__init__()
self.cache = {}
module_weights = {}
for key, value in state_dict.items():
fragments = key.split(".")
module_name = fragments[0]
weight_name = ".".join(fragments[1:])
if module_name not in module_weights:
module_weights[module_name] = {}
module_weights[module_name][weight_name] = value
modules = {}
for module_name, weights in module_weights.items():
if "conditioning1.4.weight" in weights:
depth = 3
elif weights["conditioning1.2.weight"].shape[-1] == 4:
depth = 2
else:
depth = 1
module = LLLiteModule(
name=module_name,
is_conv2d=weights["down.0.weight"].ndim == 4,
in_dim=weights["down.0.weight"].shape[1],
depth=depth,
cond_emb_dim=weights["conditioning1.0.weight"].shape[0] * 2,
mlp_dim=weights["down.0.weight"].shape[0],
)
info = module.load_state_dict(weights)
modules[module_name] = module
setattr(self, module_name, module)
if len(modules) == 1:
module.is_first = True
self.modules = modules
return
def reset(self):
self.cache = {}
return
@torch.no_grad()
def hook(self, model, cond, weight, start, end):
global all_hack
cond_image = cond * 2.0 - 1.0
for module in self.modules.values():
module.set_cond_image(cond_image)
for k, v in self.modules.items():
k = k.replace('middle_block', 'middle_blocks_0')
match = re.match("lllite_unet_(.*)_blocks_(.*)_1_transformer_blocks_(.*)_(.*)_to_(.*)", k, re.M | re.I)
assert match, 'Failed to load ControlLLLite!'
root = match.group(1)
block = match.group(2)
block_number = match.group(3)
attn_name = match.group(4)
proj_name = match.group(5)
if root == 'input':
b = model.input_blocks[int(block)][1].transformer_blocks[int(block_number)]
elif root == 'output':
b = model.output_blocks[int(block)][1].transformer_blocks[int(block_number)]
else:
b = model.middle_block[1].transformer_blocks[int(block_number)]
b = getattr(b, attn_name, None)
assert b is not None, 'Failed to load ControlLLLite!'
b = getattr(b, 'to_' + proj_name, None)
assert b is not None, 'Failed to load ControlLLLite!'
if not hasattr(b, 'lllite_list'):
b.lllite_list = []
if len(b.lllite_list) == 0:
all_hack[b] = b.forward
b.forward = self.get_hacked_forward(original_forward=b.forward, model=model, blk=b)
b.lllite_list.append((weight, start, end, v))
return
def get_hacked_forward(self, original_forward, model, blk):
@torch.no_grad()
def forward(x, **kwargs):
current_sampling_percent = getattr(model, 'current_sampling_percent', 0.5)
hackers = blk.lllite_list
hack = 0
for weight, start, end, module in hackers:
module.to(x.device)
if current_sampling_percent < start or current_sampling_percent > end:
hack = hack + 0
else:
hack = hack + module(x) * weight
x = x + hack
return original_forward(x, **kwargs)
return forward

191
scripts/controlnet_lora.py Normal file
View File

@ -0,0 +1,191 @@
import torch
from contextlib import contextmanager
from typing import Union, Tuple
_size_2_t = Union[int, Tuple[int, int]]
class LinearWithLoRA(torch.nn.Module):
def __init__(
self,
in_features: int,
out_features: int,
bias: bool = True,
device=None,
dtype=None) -> None:
super().__init__()
self.weight_module = None
self.up = None
self.down = None
self.bias = None
self.in_features = in_features
self.out_features = out_features
self.device = device
self.dtype = dtype
self.weight = None
def bind_lora(self, weight_module):
self.weight_module = [weight_module]
def unbind_lora(self):
if self.up is not None and self.down is not None: # SAI's model is weird and needs this
self.weight_module = None
def get_original_weight(self):
if self.weight_module is None:
return None
return self.weight_module[0].weight
def forward(self, x):
if self.weight is not None:
return torch.nn.functional.linear(x, self.weight.to(x),
self.bias.to(x) if self.bias is not None else None)
original_weight = self.get_original_weight()
if original_weight is None:
return None # A1111 needs first_time_calculation
if self.up is not None and self.down is not None:
weight = original_weight.to(x) + torch.mm(self.up, self.down).to(x)
else:
weight = original_weight.to(x)
return torch.nn.functional.linear(x, weight, self.bias.to(x) if self.bias is not None else None)
class Conv2dWithLoRA(torch.nn.Module):
def __init__(
self,
in_channels: int,
out_channels: int,
kernel_size: _size_2_t,
stride: _size_2_t = 1,
padding: Union[str, _size_2_t] = 0,
dilation: _size_2_t = 1,
groups: int = 1,
bias: bool = True,
padding_mode: str = 'zeros',
device=None,
dtype=None
) -> None:
super().__init__()
self.stride = stride
self.padding = padding
self.dilation = dilation
self.groups = groups
self.weight_module = None
self.bias = None
self.up = None
self.down = None
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.padding_mode = padding_mode
self.device = device
self.dtype = dtype
self.weight = None
def bind_lora(self, weight_module):
self.weight_module = [weight_module]
def unbind_lora(self):
if self.up is not None and self.down is not None: # SAI's model is weird and needs this
self.weight_module = None
def get_original_weight(self):
if self.weight_module is None:
return None
return self.weight_module[0].weight
def forward(self, x):
if self.weight is not None:
return torch.nn.functional.conv2d(x, self.weight.to(x), self.bias.to(x) if self.bias is not None else None,
self.stride, self.padding, self.dilation, self.groups)
original_weight = self.get_original_weight()
if original_weight is None:
return None # A1111 needs first_time_calculation
if self.up is not None and self.down is not None:
weight = original_weight.to(x) + torch.mm(self.up.flatten(start_dim=1), self.down.flatten(start_dim=1)).reshape(original_weight.shape).to(x)
else:
weight = original_weight.to(x)
return torch.nn.functional.conv2d(x, weight, self.bias.to(x) if self.bias is not None else None,
self.stride, self.padding, self.dilation, self.groups)
@contextmanager
def controlnet_lora_hijack():
linear, conv2d = torch.nn.Linear, torch.nn.Conv2d
torch.nn.Linear, torch.nn.Conv2d = LinearWithLoRA, Conv2dWithLoRA
try:
yield
finally:
torch.nn.Linear, torch.nn.Conv2d = linear, conv2d
def recursive_set(obj, key, value):
if obj is None:
return
if '.' in key:
k1, k2 = key.split('.', 1)
recursive_set(getattr(obj, k1, None), k2, value)
else:
setattr(obj, key, value)
def force_load_state_dict(model, state_dict):
for k in list(state_dict.keys()):
recursive_set(model, k, torch.nn.Parameter(state_dict[k]))
del state_dict[k]
return
def recursive_bind_lora(obj, key, value):
if obj is None:
return
if '.' in key:
k1, k2 = key.split('.', 1)
recursive_bind_lora(getattr(obj, k1, None), k2, value)
else:
target = getattr(obj, key, None)
if target is not None and hasattr(target, 'bind_lora'):
target.bind_lora(value)
def recursive_get(obj, key):
if obj is None:
return
if '.' in key:
k1, k2 = key.split('.', 1)
return recursive_get(getattr(obj, k1, None), k2)
else:
return getattr(obj, key, None)
def bind_control_lora(base_model, control_lora_model):
sd = base_model.state_dict()
keys = list(sd.keys())
keys = list(set([k.rsplit('.', 1)[0] for k in keys]))
module_dict = {k: recursive_get(base_model, k) for k in keys}
for k, v in module_dict.items():
recursive_bind_lora(control_lora_model, k, v)
def torch_dfs(model: torch.nn.Module):
result = [model]
for child in model.children():
result += torch_dfs(child)
return result
def unbind_control_lora(control_lora_model):
for m in torch_dfs(control_lora_model):
if hasattr(m, 'unbind_lora'):
m.unbind_lora()
return

216
scripts/controlnet_model_guess.py Normal file
View File

@ -0,0 +1,216 @@
import copy
import os
import torch
from pathlib import Path
from modules import devices
from scripts.adapter import PlugableAdapter, Adapter, StyleAdapter, Adapter_light
from scripts.controlnet_lllite import PlugableControlLLLite
from scripts.cldm import PlugableControlModel
from scripts.controlmodel_ipadapter import PlugableIPAdapter
from scripts.logging import logger
from scripts.controlnet_diffusers import convert_from_diffuser_state_dict
from scripts.controlnet_lora import controlnet_lora_hijack, force_load_state_dict
controlnet_default_config = {'adm_in_channels': None,
'in_channels': 4,
'model_channels': 320,
'num_res_blocks': 2,
'attention_resolutions': [1, 2, 4],
'transformer_depth': [1, 1, 1, 0],
'channel_mult': [1, 2, 4, 4],
'transformer_depth_middle': 1,
'use_linear_in_transformer': False,
'context_dim': 768,
"num_heads": 8,
"global_average_pooling": False}
controlnet_sdxl_config = {'num_classes': 'sequential',
'adm_in_channels': 2816,
'in_channels': 4,
'model_channels': 320,
'num_res_blocks': 2,
'attention_resolutions': [2, 4],
'transformer_depth': [0, 2, 10],
'channel_mult': [1, 2, 4],
'transformer_depth_middle': 10,
'use_linear_in_transformer': True,
'context_dim': 2048,
"num_head_channels": 64,
"global_average_pooling": False}
controlnet_sdxl_mid_config = {'num_classes': 'sequential',
'adm_in_channels': 2816,
'in_channels': 4,
'model_channels': 320,
'num_res_blocks': 2,
'attention_resolutions': [4],
'transformer_depth': [0, 0, 1],
'channel_mult': [1, 2, 4],
'transformer_depth_middle': 1,
'use_linear_in_transformer': True,
'context_dim': 2048,
"num_head_channels": 64,
"global_average_pooling": False}
controlnet_sdxl_small_config = {'num_classes': 'sequential',
'adm_in_channels': 2816,
'in_channels': 4,
'model_channels': 320,
'num_res_blocks': 2,
'attention_resolutions': [],
'transformer_depth': [0, 0, 0],
'channel_mult': [1, 2, 4],
'transformer_depth_middle': 0,
'use_linear_in_transformer': True,
"num_head_channels": 64,
'context_dim': 1,
"global_average_pooling": False}
t2i_adapter_config = {
'channels': [320, 640, 1280, 1280],
'nums_rb': 2,
'ksize': 1,
'sk': True,
'cin': 192,
'use_conv': False
}
t2i_adapter_light_config = {
'channels': [320, 640, 1280, 1280],
'nums_rb': 4,
'cin': 192,
}
t2i_adapter_style_config = {
'width': 1024,
'context_dim': 768,
'num_head': 8,
'n_layes': 3,
'num_token': 8,
}
def build_model_by_guess(state_dict, unet, model_path):
if "lora_controlnet" in state_dict:
del state_dict['lora_controlnet']
config = copy.deepcopy(controlnet_sdxl_config)
logger.info('controlnet_sdxl_config (using lora)')
config['global_average_pooling'] = False
config['hint_channels'] = int(state_dict['input_hint_block.0.weight'].shape[1])
config['use_fp16'] = devices.dtype_unet == torch.float16
with controlnet_lora_hijack():
network = PlugableControlModel(config, state_dict=None)
force_load_state_dict(network.control_model, state_dict)
network.is_control_lora = True
network.to(devices.dtype_unet)
return network
if "controlnet_cond_embedding.conv_in.weight" in state_dict:
state_dict = convert_from_diffuser_state_dict(state_dict)
model_has_shuffle_in_filename = 'shuffle' in Path(os.path.abspath(model_path)).stem.lower()
state_dict = {k.replace("control_model.", ""): v for k, v in state_dict.items()}
state_dict = {k.replace("adapter.", ""): v for k, v in state_dict.items()}
if 'input_hint_block.0.weight' in state_dict:
if 'label_emb.0.0.bias' not in state_dict:
config = copy.deepcopy(controlnet_default_config)
logger.info('controlnet_default_config')
config['global_average_pooling'] = model_has_shuffle_in_filename
config['hint_channels'] = int(state_dict['input_hint_block.0.weight'].shape[1])
config['context_dim'] = int(state_dict['input_blocks.5.1.transformer_blocks.0.attn2.to_k.weight'].shape[1])
for key in state_dict.keys():
p = state_dict[key]
if 'proj_in.weight' in key or 'proj_out.weight' in key:
if len(p.shape) == 2:
p = p[..., None, None]
state_dict[key] = p
else:
has_full_layers = 'input_blocks.8.1.transformer_blocks.9.norm3.weight' in state_dict
has_mid_layers = 'input_blocks.8.1.transformer_blocks.0.norm3.weight' in state_dict
if has_full_layers:
config = copy.deepcopy(controlnet_sdxl_config)
logger.info('controlnet_sdxl_config')
elif has_mid_layers:
config = copy.deepcopy(controlnet_sdxl_mid_config)
logger.info('controlnet_sdxl_mid_config')
else:
config = copy.deepcopy(controlnet_sdxl_small_config)
logger.info('controlnet_sdxl_small_config')
config['global_average_pooling'] = False
config['hint_channels'] = int(state_dict['input_hint_block.0.weight'].shape[1])
if 'difference' in state_dict and unet is not None:
unet_state_dict = unet.state_dict()
unet_state_dict_keys = unet_state_dict.keys()
final_state_dict = {}
for key in state_dict.keys():
p = state_dict[key]
if key in unet_state_dict_keys:
p_new = p + unet_state_dict[key].clone().cpu()
else:
p_new = p
final_state_dict[key] = p_new
state_dict = final_state_dict
config['use_fp16'] = devices.dtype_unet == torch.float16
network = PlugableControlModel(config, state_dict)
network.to(devices.dtype_unet)
return network
if 'conv_in.weight' in state_dict:
logger.info('t2i_adapter_config')
cin = int(state_dict['conv_in.weight'].shape[1])
channel = int(state_dict['conv_in.weight'].shape[0])
ksize = 1
down_opts = tuple(filter(lambda item: item.endswith("down_opt.op.weight"), state_dict))
use_conv = len(down_opts) > 0
is_sdxl = (cin % 256) == 0
adapter = Adapter(
cin=cin,
channels=[channel, channel*2, channel*4, channel*4],
nums_rb=2,
ksize=ksize,
sk=True,
use_conv=use_conv,
is_sdxl=is_sdxl
).cpu()
adapter.load_state_dict(state_dict, strict=False)
network = PlugableAdapter(adapter)
return network
if 'style_embedding' in state_dict:
config = copy.deepcopy(t2i_adapter_style_config)
logger.info('t2i_adapter_style_config')
adapter = StyleAdapter(**config).cpu()
adapter.load_state_dict(state_dict, strict=False)
network = PlugableAdapter(adapter)
return network
if 'body.0.in_conv.weight' in state_dict:
config = copy.deepcopy(t2i_adapter_light_config)
logger.info('t2i_adapter_light_config')
config['cin'] = int(state_dict['body.0.in_conv.weight'].shape[1])
adapter = Adapter_light(**config).cpu()
adapter.load_state_dict(state_dict, strict=False)
network = PlugableAdapter(adapter)
return network
if 'ip_adapter' in state_dict:
plus = "latents" in state_dict["image_proj"]
if plus:
channel = int(state_dict['image_proj']['proj_in.weight'].shape[1])
else:
channel = int(state_dict['image_proj']['proj.weight'].shape[1])
network = PlugableIPAdapter(state_dict, channel, plus)
network.to('cpu')
return network
if any('lllite' in k for k in state_dict.keys()):
network = PlugableControlLLLite(state_dict)
network.to('cpu')
return network
raise '[ControlNet Error] Cannot recognize the ControlModel!'

22
scripts/controlnet_ui/controlnet_ui_group.py
View File

@ -14,6 +14,7 @@ from scripts import (
)
from scripts.processor import (
preprocessor_sliders_config,
no_control_mode_preprocessors,
flag_preprocessor_resolution,
model_free_preprocessors,
preprocessor_filters,
@ -566,6 +567,8 @@ class ControlNetUiGroup(object):
0, self.prevent_next_n_slider_value_update - 1
)
grs += [gr.update(visible=module not in no_control_mode_preprocessors)]
return grs
inputs = [
@ -579,6 +582,7 @@ class ControlNetUiGroup(object):
self.advanced,
self.model,
self.refresh_models,
self.control_mode
]
self.module.change(build_sliders, inputs=inputs, outputs=outputs)
self.pixel_perfect.change(build_sliders, inputs=inputs, outputs=outputs)
@ -682,24 +686,14 @@ class ControlNetUiGroup(object):
else None,
)
if "clip" in module:
result = processor.clip_vision_visualization(result)
if not is_image:
result = img
is_image = True
if is_image:
result = external_code.visualize_inpaint_mask(result)
return (
# Update to `generated_image`
gr.update(value=result, visible=True, interactive=False),
# preprocessor_preview
gr.update(value=True),
# openpose editor
*self.openpose_editor.update(json_acceptor.value),
)
result = external_code.visualize_inpaint_mask(result)
return (
# Update to `generated_image`
gr.update(value=None, visible=True),
gr.update(value=result, visible=True, interactive=False),
# preprocessor_preview
gr.update(value=True),
# openpose editor

2
scripts/controlnet_version.py
View File

@ -1,4 +1,4 @@
version_flag = 'v1.1.313'
version_flag = 'v1.1.400'
from scripts.logging import logger

23
scripts/global_state.py
View File

@ -63,7 +63,11 @@ cn_preprocessor_modules = {
"openpose_faceonly": functools.partial(g_openpose_model.run_model, include_body=False, include_hand=False, include_face=True),
"openpose_full": functools.partial(g_openpose_model.run_model, include_body=True, include_hand=True, include_face=True),
"dw_openpose_full": functools.partial(g_openpose_model.run_model, include_body=True, include_hand=True, include_face=True, use_dw_pose=True),
"clip_vision": clip,
"clip_vision": functools.partial(clip, config='clip_vitl'),
"revision_clipvision": functools.partial(clip, config='clip_g'),
"revision_ignore_prompt": functools.partial(clip, config='clip_g'),
"ip-adapter_clip_sd15": functools.partial(clip, config='clip_h'),
"ip-adapter_clip_sdxl": functools.partial(clip, config='clip_g'),
"color": color,
"pidinet": pidinet,
"pidinet_safe": pidinet_safe,
@ -93,13 +97,19 @@ cn_preprocessor_modules = {
"inpaint_only+lama": lama_inpaint,
"tile_colorfix": identity,
"tile_colorfix+sharp": identity,
"recolor_luminance": recolor_luminance,
"recolor_intensity": recolor_intensity,
}
cn_preprocessor_unloadable = {
"hed": unload_hed,
"fake_scribble": unload_hed,
"mlsd": unload_mlsd,
"clip": unload_clip,
"clip_vision": functools.partial(unload_clip, config='clip_vitl'),
"revision_clipvision": functools.partial(unload_clip, config='clip_g'),
"revision_ignore_prompt": functools.partial(unload_clip, config='clip_g'),
"ip-adapter_clip_sd15": functools.partial(unload_clip, config='clip_h'),
"ip-adapter_clip_sdxl": functools.partial(unload_clip, config='clip_g'),
"depth": unload_midas,
"depth_leres": unload_leres,
"normal_map": unload_midas,
@ -148,18 +158,19 @@ ui_preprocessor_keys += sorted([preprocessor_aliases.get(k, k)
reverse_preprocessor_aliases = {preprocessor_aliases[k]: k for k in preprocessor_aliases.keys()}
def get_module_basename(module: Optional[str]) -> str:
if module is None:
module = 'none'
return reverse_preprocessor_aliases.get(module, module)
default_conf = os.path.join("models", "cldm_v15.yaml")
default_conf_adapter = os.path.join("models", "t2iadapter_sketch_sd14v1.yaml")
default_detectedmap_dir = os.path.join("detected_maps")
script_dir = scripts.basedir()
os.makedirs(cn_models_dir, exist_ok=True)
def traverse_all_files(curr_path, model_list):
f_list = [
(os.path.join(curr_path, entry.name), entry.stat())
@ -241,14 +252,14 @@ def select_control_type(control_type: str) -> Tuple[List[str], List[str], str, s
filtered_preprocessor_list = [
x
for x in preprocessor_list
if pattern in x.lower() or x.lower() == "none"
if pattern in x.lower() or any(a in x.lower() for a in preprocessor_filters_aliases.get(pattern, [])) or x.lower() == "none"
]
if pattern in ["canny", "lineart", "scribble", "mlsd"]:
filtered_preprocessor_list += [
x for x in preprocessor_list if "invert" in x.lower()
]
filtered_model_list = [
x for x in model_list if pattern in x.lower() or x.lower() == "none"
x for x in model_list if pattern in x.lower() or any(a in x.lower() for a in preprocessor_filters_aliases.get(pattern, [])) or x.lower() == "none"
]
if default_option not in filtered_preprocessor_list:
default_option = filtered_preprocessor_list[0]

320
scripts/hook.py
View File

@ -3,9 +3,10 @@ import einops
import hashlib
import numpy as np
import torch.nn as nn
from functools import partial
import modules.processing
from enum import Enum
from scripts.logging import logger
from modules import devices, lowvram, shared, scripts
@ -21,6 +22,13 @@ from modules.prompt_parser import MulticondLearnedConditioning, ComposableSchedu
from modules.processing import StableDiffusionProcessing
try:
from sgm.modules.attention import BasicTransformerBlock as BasicTransformerBlockSGM
except:
print('Webui version too old!')
BasicTransformerBlockSGM = BasicTransformerBlock
POSITIVE_MARK_TOKEN = 1024
NEGATIVE_MARK_TOKEN = - POSITIVE_MARK_TOKEN
MARK_EPS = 1e-3
@ -45,12 +53,20 @@ def mark_prompt_context(x, positive):
x.schedules = mark_prompt_context(x.schedules, positive)
return x
if isinstance(x, ScheduledPromptConditioning):
cond = x.cond
if prompt_context_is_marked(cond):
return x
mark = POSITIVE_MARK_TOKEN if positive else NEGATIVE_MARK_TOKEN
cond = torch.cat([torch.zeros_like(cond)[:1] + mark, cond], dim=0)
return ScheduledPromptConditioning(end_at_step=x.end_at_step, cond=cond)
if isinstance(x.cond, dict):
cond = x.cond['crossattn']
if prompt_context_is_marked(cond):
return x
mark = POSITIVE_MARK_TOKEN if positive else NEGATIVE_MARK_TOKEN
cond = torch.cat([torch.zeros_like(cond)[:1] + mark, cond], dim=0)
return ScheduledPromptConditioning(end_at_step=x.end_at_step, cond=dict(crossattn=cond, vector=x.cond['vector']))
else:
cond = x.cond
if prompt_context_is_marked(cond):
return x
mark = POSITIVE_MARK_TOKEN if positive else NEGATIVE_MARK_TOKEN
cond = torch.cat([torch.zeros_like(cond)[:1] + mark, cond], dim=0)
return ScheduledPromptConditioning(end_at_step=x.end_at_step, cond=cond)
return x
@ -91,27 +107,22 @@ def unmark_prompt_context(x):
return mark_batch, uc_indices, context
def create_random_tensors_hacked(*args, **kwargs):
result = modules.processing.create_random_tensors_original(*args, **kwargs)
p = kwargs.get('p', None)
if p is None:
return result
controlnet_initial_noise_modifier = getattr(p, 'controlnet_initial_noise_modifier', None)
if controlnet_initial_noise_modifier is not None:
x0 = controlnet_initial_noise_modifier
class HackedImageRNG:
def __init__(self, rng, noise_modifier, sd_model):
self.rng = rng
self.noise_modifier = noise_modifier
self.sd_model = sd_model
def next(self):
result = self.rng.next()
x0 = self.noise_modifier
if result.shape[2] != x0.shape[2] or result.shape[3] != x0.shape[3]:
return result
x0 = x0.to(result.dtype).to(result.device)
ts = torch.tensor([p.sd_model.num_timesteps - 1] * result.shape[0]).long().to(result.device)
result = p.sd_model.q_sample(x0, ts, result)
ts = torch.tensor([999] * result.shape[0]).long().to(result.device)
result = predict_q_sample(self.sd_model, x0, ts, result)
logger.info(f'[ControlNet] Initial noise hack applied to {result.shape}.')
return result
if getattr(modules.processing, 'create_random_tensors_original', None) is None:
modules.processing.create_random_tensors_original = modules.processing.create_random_tensors
modules.processing.create_random_tensors = create_random_tensors_hacked
return result
class ControlModelType(Enum):
@ -129,6 +140,9 @@ class ControlModelType(Enum):
StableSR = "StableSR, Jianyi Wang"
PromptDiffusion = "PromptDiffusion, Zhendong Wang"
ControlLoRA = "ControlLoRA, Wu Hecong"
ReVision = "ReVision, Stability"
IPAdapter = "IPAdapter, Hu Ye"
Controlllite = "Controlllite, Kohya"
# Written by Lvmin
@ -252,12 +266,41 @@ def torch_dfs(model: torch.nn.Module):
return result
def register_schedule(self):
linear_start = 0.00085
linear_end = 0.0120
num_timesteps = 1000
betas = (torch.linspace(linear_start ** 0.5, linear_end ** 0.5, num_timesteps, dtype=torch.float64) ** 2.0).numpy()
alphas = 1. - betas
alphas_cumprod = np.cumprod(alphas, axis=0)
alphas_cumprod_prev = np.append(1., alphas_cumprod[:-1])
to_torch = partial(torch.tensor, dtype=torch.float32)
setattr(self, 'betas', to_torch(betas))
# setattr(self, 'alphas_cumprod', to_torch(alphas_cumprod)) # a1111 already has this
setattr(self, 'alphas_cumprod_prev', to_torch(alphas_cumprod_prev))
setattr(self, 'sqrt_alphas_cumprod', to_torch(np.sqrt(alphas_cumprod)))
setattr(self, 'sqrt_one_minus_alphas_cumprod', to_torch(np.sqrt(1. - alphas_cumprod)))
setattr(self, 'log_one_minus_alphas_cumprod', to_torch(np.log(1. - alphas_cumprod)))
setattr(self, 'sqrt_recip_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod)))
setattr(self, 'sqrt_recipm1_alphas_cumprod', to_torch(np.sqrt(1. / alphas_cumprod - 1)))
def predict_q_sample(ldm, x_start, t, noise=None):
if noise is None:
noise = torch.randn_like(x_start)
return extract_into_tensor(ldm.sqrt_alphas_cumprod.to(x_start), t, x_start.shape) * x_start + extract_into_tensor(ldm.sqrt_one_minus_alphas_cumprod.to(x_start), t, x_start.shape) * noise
def predict_start_from_noise(ldm, x_t, t, noise):
return extract_into_tensor(ldm.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - extract_into_tensor(ldm.sqrt_recipm1_alphas_cumprod, t, x_t.shape) * noise
return extract_into_tensor(ldm.sqrt_recip_alphas_cumprod.to(x_t), t, x_t.shape) * x_t - extract_into_tensor(ldm.sqrt_recipm1_alphas_cumprod.to(x_t), t, x_t.shape) * noise
def predict_noise_from_start(ldm, x_t, t, x0):
return (extract_into_tensor(ldm.sqrt_recip_alphas_cumprod, t, x_t.shape) * x_t - x0) / extract_into_tensor(ldm.sqrt_recipm1_alphas_cumprod, t, x_t.shape)
return (extract_into_tensor(ldm.sqrt_recip_alphas_cumprod.to(x_t), t, x_t.shape) * x_t - x0) / extract_into_tensor(ldm.sqrt_recipm1_alphas_cumprod.to(x_t), t, x_t.shape)
def blur(x, k):
@ -298,6 +341,7 @@ class UnetHook(nn.Module):
self.gn_auto_machine_weight = 1.0
self.current_style_fidelity = 0.0
self.current_uc_indices = None
self.global_revision = None
@staticmethod
def call_vae_using_process(p, x, batch_size=None, mask=None):
@ -317,6 +361,15 @@ class UnetHook(nn.Module):
with devices.autocast():
vae_output = p.sd_model.encode_first_stage(x)
vae_output = p.sd_model.get_first_stage_encoding(vae_output)
if torch.all(torch.isnan(vae_output)).item():
logger.info(f'ControlNet find Nans in the VAE encoding. \n '
f'Now ControlNet will automatically retry.\n '
f'To always start with 32-bit VAE, use --no-half-vae commandline flag.')
devices.dtype_vae = torch.float32
x = x.to(devices.dtype_vae)
p.sd_model.first_stage_model.to(devices.dtype_vae)
vae_output = p.sd_model.encode_first_stage(x)
vae_output = p.sd_model.get_first_stage_encoding(vae_output)
vae_cache.set(x, vae_output)
logger.info(f'ControlNet used {str(devices.dtype_vae)} VAE to encode {vae_output.shape}.')
latent = vae_output
@ -332,12 +385,16 @@ class UnetHook(nn.Module):
for param in self.control_params:
current_sampling_percent = (x.sampling_step / x.total_sampling_steps)
param.guidance_stopped = current_sampling_percent < param.start_guidance_percent or current_sampling_percent > param.stop_guidance_percent
if self.model is not None:
self.model.current_sampling_percent = current_sampling_percent
def hook(self, model, sd_ldm, control_params, process):
self.model = model
self.sd_ldm = sd_ldm
self.control_params = control_params
model_is_sdxl = getattr(self.sd_ldm, 'is_sdxl', False)
outer = self
def process_sample(*args, **kwargs):
@ -354,17 +411,28 @@ class UnetHook(nn.Module):
mark_prompt_context(getattr(process, 'hr_uc', []), positive=False)
return process.sample_before_CN_hack(*args, **kwargs)
def forward(self, x, timesteps=None, context=None, **kwargs):
total_controlnet_embedding = [0.0] * 13
def forward(self, x, timesteps=None, context=None, y=None, **kwargs):
is_sdxl = y is not None and model_is_sdxl
total_t2i_adapter_embedding = [0.0] * 4
if is_sdxl:
total_controlnet_embedding = [0.0] * 10
else:
total_controlnet_embedding = [0.0] * 13
require_inpaint_hijack = False
is_in_high_res_fix = False
batch_size = int(x.shape[0])
# Handle cond-uncond marker
cond_mark, outer.current_uc_indices, context = unmark_prompt_context(context)
outer.model.cond_mark = cond_mark
# logger.info(str(cond_mark[:, 0, 0, 0].detach().cpu().numpy().tolist()) + ' - ' + str(outer.current_uc_indices))
# Revision
if is_sdxl and isinstance(outer.global_revision, torch.Tensor):
y[:, :1280] = outer.global_revision * cond_mark[:, :, 0, 0]
if any('ignore_prompt' in param.preprocessor['name'] for param in outer.control_params):
context = torch.zeros_like(context)
# High-res fix
for param in outer.control_params:
# select which hint_cond to use
@ -404,6 +472,19 @@ class UnetHook(nn.Module):
continue
param.used_hint_cond_latent = outer.call_vae_using_process(process, param.used_hint_cond, batch_size=batch_size)
# vram
for param in outer.control_params:
if getattr(param.control_model, 'disable_memory_management', False):
continue
if param.control_model is not None:
if outer.lowvram and is_sdxl and hasattr(param.control_model, 'aggressive_lowvram'):
param.control_model.aggressive_lowvram()
elif hasattr(param.control_model, 'fullvram'):
param.control_model.fullvram()
elif hasattr(param.control_model, 'to'):
param.control_model.to(devices.get_device_for("controlnet"))
# handle prompt token control
for param in outer.control_params:
if no_high_res_control:
@ -415,7 +496,6 @@ class UnetHook(nn.Module):
if param.control_model_type not in [ControlModelType.T2I_StyleAdapter]:
continue
param.control_model.to(devices.get_device_for("controlnet"))
control = param.control_model(x=x, hint=param.used_hint_cond, timesteps=timesteps, context=context)
control = torch.cat([control.clone() for _ in range(batch_size)], dim=0)
control *= param.weight
@ -433,7 +513,6 @@ class UnetHook(nn.Module):
if param.control_model_type not in [ControlModelType.ControlNet, ControlModelType.T2I_Adapter]:
continue
param.control_model.to(devices.get_device_for("controlnet"))
# inpaint model workaround
x_in = x
control_model = param.control_model.control_model
@ -455,11 +534,12 @@ class UnetHook(nn.Module):
m = (m > 0.5).float()
hint = c * (1 - m) - m
control = param.control_model(x=x_in, hint=hint, timesteps=timesteps, context=context)
control_scales = ([param.weight] * 13)
control = param.control_model(x=x_in, hint=hint, timesteps=timesteps, context=context, y=y)
if outer.lowvram:
param.control_model.to("cpu")
if is_sdxl:
control_scales = [param.weight] * 10
else:
control_scales = [param.weight] * 13
if param.cfg_injection or param.global_average_pooling:
if param.control_model_type == ControlModelType.T2I_Adapter:
@ -484,6 +564,9 @@ class UnetHook(nn.Module):
elif param.control_model_type == ControlModelType.ControlNet:
control_scales = [param.weight * (0.825 ** float(12 - i)) for i in range(13)]
if is_sdxl and param.control_model_type == ControlModelType.ControlNet:
control_scales = control_scales[:10]
if param.advanced_weighting is not None:
control_scales = param.advanced_weighting
@ -502,6 +585,8 @@ class UnetHook(nn.Module):
# Replace x_t to support inpaint models
for param in outer.control_params:
if not isinstance(param.used_hint_cond, torch.Tensor):
continue
if param.used_hint_cond.shape[1] != 4:
continue
if x.shape[1] != 9:
@ -518,8 +603,14 @@ class UnetHook(nn.Module):
param.used_hint_inpaint_hijack.to(x.dtype).to(x.device)
x = torch.cat([x[:, :4, :, :], param.used_hint_inpaint_hijack], dim=1)
# vram
for param in outer.control_params:
if param.control_model is not None:
if outer.lowvram:
param.control_model.to('cpu')
# A1111 fix for medvram.
if shared.cmd_opts.medvram:
if shared.cmd_opts.medvram or (getattr(shared.cmd_opts, 'medvram_sdxl', False) and is_sdxl):
try:
# Trigger the register_forward_pre_hook
outer.sd_ldm.model()
@ -549,7 +640,7 @@ class UnetHook(nn.Module):
if param.control_model_type not in [ControlModelType.AttentionInjection]:
continue
ref_xt = outer.sd_ldm.q_sample(param.used_hint_cond_latent, torch.round(timesteps.float()).long())
ref_xt = predict_q_sample(outer.sd_ldm, param.used_hint_cond_latent, torch.round(timesteps.float()).long())
# Inpaint Hijack
if x.shape[1] == 9:
@ -562,6 +653,12 @@ class UnetHook(nn.Module):
outer.current_style_fidelity = float(param.preprocessor['threshold_a'])
outer.current_style_fidelity = max(0.0, min(1.0, outer.current_style_fidelity))
if is_sdxl:
# sdxl's attention hacking is highly unstable.
# We have no other methods but to reduce the style_fidelity a bit.
# By default, 0.5 ** 3.0 = 0.125
outer.current_style_fidelity = outer.current_style_fidelity ** 3.0
if param.cfg_injection:
outer.current_style_fidelity = 1.0
elif param.soft_injection or is_in_high_res_fix:
@ -577,11 +674,19 @@ class UnetHook(nn.Module):
outer.gn_auto_machine = AutoMachine.Write
outer.gn_auto_machine_weight = param.weight
outer.original_forward(
x=ref_xt.to(devices.dtype_unet),
timesteps=timesteps.to(devices.dtype_unet),
context=context.to(devices.dtype_unet)
)
if is_sdxl:
outer.original_forward(
x=ref_xt.to(devices.dtype_unet),
timesteps=timesteps.to(devices.dtype_unet),
context=context.to(devices.dtype_unet),
y=y
)
else:
outer.original_forward(
x=ref_xt.to(devices.dtype_unet),
timesteps=timesteps.to(devices.dtype_unet),
context=context.to(devices.dtype_unet)
)
outer.attention_auto_machine = AutoMachine.Read
outer.gn_auto_machine = AutoMachine.Read
@ -591,11 +696,18 @@ class UnetHook(nn.Module):
with th.no_grad():
t_emb = cond_cast_unet(timestep_embedding(timesteps, self.model_channels, repeat_only=False))
emb = self.time_embed(t_emb)
h = x.type(self.dtype)
if is_sdxl:
assert y.shape[0] == x.shape[0]
emb = emb + self.label_emb(y)
h = x
for i, module in enumerate(self.input_blocks):
h = module(h, emb, context)
if (i + 1) % 3 == 0:
t2i_injection = [3, 5, 8] if is_sdxl else [2, 5, 8, 11]
if i in t2i_injection:
h = aligned_adding(h, total_t2i_adapter_embedding.pop(0), require_inpaint_hijack)
hs.append(h)
@ -604,6 +716,9 @@ class UnetHook(nn.Module):
# U-Net Middle Block
h = aligned_adding(h, total_controlnet_embedding.pop(), require_inpaint_hijack)
if len(total_t2i_adapter_embedding) > 0 and is_sdxl:
h = aligned_adding(h, total_t2i_adapter_embedding.pop(0), require_inpaint_hijack)
# U-Net Decoder
for i, module in enumerate(self.output_blocks):
h = th.cat([h, aligned_adding(hs.pop(), total_controlnet_embedding.pop(), require_inpaint_hijack)], dim=1)
@ -668,18 +783,23 @@ class UnetHook(nn.Module):
return h
def move_all_control_model_to_cpu():
for param in getattr(outer, 'control_params', []):
if isinstance(param.control_model, torch.nn.Module):
param.control_model.to("cpu")
def forward_webui(*args, **kwargs):
# webui will handle other compoments
try:
if shared.cmd_opts.lowvram:
lowvram.send_everything_to_cpu()
return forward(*args, **kwargs)
except Exception as e:
move_all_control_model_to_cpu()
raise e
finally:
if self.lowvram:
for param in self.control_params:
if isinstance(param.control_model, torch.nn.Module):
param.control_model.to("cpu")
if outer.lowvram:
move_all_control_model_to_cpu()
def hacked_basic_transformer_inner_forward(self, x, context=None):
x_norm1 = self.norm1(x)
@ -716,7 +836,7 @@ class UnetHook(nn.Module):
def hacked_group_norm_forward(self, *args, **kwargs):
eps = 1e-6
x = self.original_forward(*args, **kwargs)
x = self.original_forward_cn_hijack(*args, **kwargs)
y = None
if outer.gn_auto_machine == AutoMachine.Write:
if outer.gn_auto_machine_weight > self.gn_weight:
@ -752,56 +872,78 @@ class UnetHook(nn.Module):
outer.original_forward = model.forward
model.forward = forward_webui.__get__(model, UNetModel)
if model_is_sdxl:
register_schedule(sd_ldm)
need_attention_hijack = False
for param in outer.control_params:
if param.control_model_type in [ControlModelType.AttentionInjection]:
need_attention_hijack = True
all_modules = torch_dfs(model)
attn_modules = [module for module in all_modules if isinstance(module, BasicTransformerBlock)]
attn_modules = sorted(attn_modules, key=lambda x: - x.norm1.normalized_shape[0])
if need_attention_hijack:
attn_modules = [module for module in all_modules if isinstance(module, BasicTransformerBlock) or isinstance(module, BasicTransformerBlockSGM)]
attn_modules = sorted(attn_modules, key=lambda x: - x.norm1.normalized_shape[0])
for i, module in enumerate(attn_modules):
if getattr(module, '_original_inner_forward', None) is None:
module._original_inner_forward = module._forward
module._forward = hacked_basic_transformer_inner_forward.__get__(module, BasicTransformerBlock)
module.bank = []
module.style_cfgs = []
module.attn_weight = float(i) / float(len(attn_modules))
for i, module in enumerate(attn_modules):
if getattr(module, '_original_inner_forward_cn_hijack', None) is None:
module._original_inner_forward_cn_hijack = module._forward
module._forward = hacked_basic_transformer_inner_forward.__get__(module, BasicTransformerBlock)
module.bank = []
module.style_cfgs = []
module.attn_weight = float(i) / float(len(attn_modules))
gn_modules = [model.middle_block]
model.middle_block.gn_weight = 0
gn_modules = [model.middle_block]
model.middle_block.gn_weight = 0
input_block_indices = [4, 5, 7, 8, 10, 11]
for w, i in enumerate(input_block_indices):
module = model.input_blocks[i]
module.gn_weight = 1.0 - float(w) / float(len(input_block_indices))
gn_modules.append(module)
if model_is_sdxl:
input_block_indices = [4, 5, 7, 8]
output_block_indices = [0, 1, 2, 3, 4, 5]
else:
input_block_indices = [4, 5, 7, 8, 10, 11]
output_block_indices = [0, 1, 2, 3, 4, 5, 6, 7]
output_block_indices = [0, 1, 2, 3, 4, 5, 6, 7]
for w, i in enumerate(output_block_indices):
module = model.output_blocks[i]
module.gn_weight = float(w) / float(len(output_block_indices))
gn_modules.append(module)
for w, i in enumerate(input_block_indices):
module = model.input_blocks[i]
module.gn_weight = 1.0 - float(w) / float(len(input_block_indices))
gn_modules.append(module)
for i, module in enumerate(gn_modules):
if getattr(module, 'original_forward', None) is None:
module.original_forward = module.forward
module.forward = hacked_group_norm_forward.__get__(module, torch.nn.Module)
module.mean_bank = []
module.var_bank = []
module.style_cfgs = []
module.gn_weight *= 2
for w, i in enumerate(output_block_indices):
module = model.output_blocks[i]
module.gn_weight = float(w) / float(len(output_block_indices))
gn_modules.append(module)
outer.attn_module_list = attn_modules
outer.gn_module_list = gn_modules
for i, module in enumerate(gn_modules):
if getattr(module, 'original_forward_cn_hijack', None) is None:
module.original_forward_cn_hijack = module.forward
module.forward = hacked_group_norm_forward.__get__(module, torch.nn.Module)
module.mean_bank = []
module.var_bank = []
module.style_cfgs = []
module.gn_weight *= 2
outer.attn_module_list = attn_modules
outer.gn_module_list = gn_modules
else:
for module in enumerate(all_modules):
_original_inner_forward_cn_hijack = getattr(module, '_original_inner_forward_cn_hijack', None)
original_forward_cn_hijack = getattr(module, 'original_forward_cn_hijack', None)
if _original_inner_forward_cn_hijack is not None:
module._forward = _original_inner_forward_cn_hijack
if original_forward_cn_hijack is not None:
module.forward = original_forward_cn_hijack
outer.attn_module_list = []
outer.gn_module_list = []
scripts.script_callbacks.on_cfg_denoiser(self.guidance_schedule_handler)
def restore(self, model):
def restore(self):
scripts.script_callbacks.remove_callbacks_for_function(self.guidance_schedule_handler)
if hasattr(self, "control_params"):
del self.control_params
self.control_params = None
if not hasattr(model, "_original_forward"):
# no such handle, ignore
return
model.forward = model._original_forward
del model._original_forward
if self.model is not None:
if hasattr(self.model, "_original_forward"):
self.model.forward = self.model._original_forward
del self.model._original_forward

4
scripts/movie2movie.py
View File

@ -83,7 +83,7 @@ class Script(scripts.Script):
# The returned values are passed to the run method as parameters.
ctrls_group = ()
max_models = opts.data.get("control_net_max_models_num", 1)
max_models = opts.data.get("control_net_unit_count", 3)
with gr.Group():
with gr.Accordion("ControlNet-M2M", open = False):
@ -109,7 +109,7 @@ class Script(scripts.Script):
# to be used in processing. The return value should be a Processed object, which is
# what is returned by the process_images method.
contents_num = opts.data.get("control_net_max_models_num", 1)
contents_num = opts.data.get("control_net_unit_count", 3)
arg_num = 3
item_list = []
video_list = []

105
scripts/processor.py
View File

@ -336,30 +336,28 @@ def unload_pidinet():
unload_pid_model()
clip_encoder = None
clip_encoder = {
'clip_g': None,
'clip_h': None,
'clip_vitl': None,
}
def clip(img, res=512, **kwargs):
def clip(img, res=512, config='clip_vitl', **kwargs):
img = HWC3(img)
global clip_encoder
if clip_encoder is None:
from annotator.clip import apply_clip
clip_encoder = apply_clip
result = clip_encoder(img)
if clip_encoder[config] is None:
from annotator.clipvision import ClipVisionDetector
clip_encoder[config] = ClipVisionDetector(config)
result = clip_encoder[config](img)
return result, False
def clip_vision_visualization(x):
x = x.detach().cpu().numpy()[0]
x = np.ascontiguousarray(x).copy()
return np.ndarray((x.shape[0] * 4, x.shape[1]), dtype="uint8", buffer=x.tobytes())
def unload_clip():
def unload_clip(config='clip_vitl'):
global clip_encoder
if clip_encoder is not None:
from annotator.clip import unload_clip_model
unload_clip_model()
if clip_encoder[config] is not None:
clip_encoder[config].unload_model()
clip_encoder[config] = None
model_color = None
@ -597,10 +595,39 @@ def shuffle(img, res=512, **kwargs):
return result, True
def recolor_luminance(img, res=512, thr_a=1.0, **kwargs):
result = cv2.cvtColor(HWC3(img), cv2.COLOR_BGR2LAB)
result = result[:, :, 0].astype(np.float32) / 255.0
result = result ** thr_a
result = (result * 255.0).clip(0, 255).astype(np.uint8)
result = cv2.cvtColor(result, cv2.COLOR_GRAY2RGB)
return result, True
def recolor_intensity(img, res=512, thr_a=1.0, **kwargs):
result = cv2.cvtColor(HWC3(img), cv2.COLOR_BGR2HSV)
result = result[:, :, 2].astype(np.float32) / 255.0
result = result ** thr_a
result = (result * 255.0).clip(0, 255).astype(np.uint8)
result = cv2.cvtColor(result, cv2.COLOR_GRAY2RGB)
return result, True
model_free_preprocessors = [
"reference_only",
"reference_adain",
"reference_adain+attn"
"reference_adain+attn",
"revision_clipvision",
"revision_ignore_prompt"
]
no_control_mode_preprocessors = [
"revision_clipvision",
"revision_ignore_prompt",
"clip_vision",
"ip-adapter_clip_sd15",
"ip-adapter_clip_sdxl",
"t2ia_style_clipvision"
]
flag_preprocessor_resolution = "Preprocessor Resolution"
@ -608,6 +635,8 @@ preprocessor_sliders_config = {
"none": [],
"inpaint": [],
"inpaint_only": [],
"revision_clipvision": [],
"revision_ignore_prompt": [],
"canny": [
{
"name": flag_preprocessor_resolution,
@ -900,23 +929,55 @@ preprocessor_sliders_config = {
"step": 0.01
}
],
"recolor_luminance": [
None,
{
"name": "Gamma Correction",
"value": 1.0,
"min": 0.1,
"max": 2.0,
"step": 0.001
}
],
"recolor_intensity": [
None,
{
"name": "Gamma Correction",
"value": 1.0,
"min": 0.1,
"max": 2.0,
"step": 0.001
}
],
}
preprocessor_filters = {
"All": "none",
"Canny": "canny",
"Depth": "depth_midas",
"Normal": "normal_bae",
"NormalMap": "normal_bae",
"OpenPose": "openpose_full",
"MLSD": "mlsd",
"Lineart": "lineart_standard (from white bg & black line)",
"SoftEdge": "softedge_pidinet",
"Scribble": "scribble_pidinet",
"Seg": "seg_ofade20k",
"Scribble/Sketch": "scribble_pidinet",
"Segmentation": "seg_ofade20k",
"Shuffle": "shuffle",
"Tile": "tile_resample",
"Inpaint": "inpaint_only",
"IP2P": "none",
"InstructP2P": "none",
"Reference": "reference_only",
"T2IA": "none",
"Recolor": "recolor_luminance",
"Revision": "revision_clipvision",
"T2I-Adapter": "none",
"IP-Adapter": "ip-adapter_clip_sd15",
}
preprocessor_filters_aliases = {
'instructp2p': ['ip2p'],
'segmentation': ['seg'],
'normalmap': ['normal'],
't2i-adapter': ['t2i_adapter', 't2iadapter', 't2ia'],
'ip-adapter': ['ip_adapter', 'ipadapter'],
'scribble/sketch': ['scribble', 'sketch']
} # must use all lower texts

10
scripts/utils.py
View File

@ -4,24 +4,20 @@ import functools
import time
import base64
import numpy as np
import gradio as gr
import safetensors.torch
import logging
from typing import Any, Callable, Dict
from modules.safe import unsafe_torch_load
from scripts.logging import logger
def load_state_dict(ckpt_path, location="cpu"):
_, extension = os.path.splitext(ckpt_path)
if extension.lower() == ".safetensors":
import safetensors.torch
state_dict = safetensors.torch.load_file(ckpt_path, device=location)
else:
state_dict = get_state_dict(
torch.load(ckpt_path, map_location=torch.device(location))
)
state_dict = unsafe_torch_load(ckpt_path, map_location=torch.device(location))
state_dict = get_state_dict(state_dict)
logger.info(f"Loaded state_dict from [{ckpt_path}]")
return state_dict

6
tests/external_code_api/external_code_test.py
View File

@ -26,13 +26,13 @@ class TestExternalCodeWorking(unittest.TestCase):
self.scripts.alwayson_scripts = [self.cn_script]
self.script_args = [None] * self.cn_script.args_from
self.initial_max_models = shared.opts.data.get("control_net_max_models_num", 1)
shared.opts.data.update(control_net_max_models_num=self.max_models)
self.initial_max_models = shared.opts.data.get("control_net_unit_count", 3)
shared.opts.data.update(control_net_unit_count=self.max_models)
self.extra_models = 0
def tearDown(self):
shared.opts.data.update(control_net_max_models_num=self.initial_max_models)
shared.opts.data.update(control_net_unit_count=self.initial_max_models)
def get_expected_args_to(self):
args_len = max(self.max_models, len(self.cn_units))