1580158
/
sd-webui-stablesr
mirror of https://github.com/pkuliyi2015/sd-webui-stablesr.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

initial commit

pull/12/head
pkuliyi2015 2023-05-20 13:39:33 +00:00
commit 97bc1ff0b3
12 changed files with 1711 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

7
.gitignore vendored Normal file
View File

@ -0,0 +1,7 @@
# meta
.vscode/
__pycache__/
.DS_Store
# settings
models/

35
LICENSE Normal file
View File

@ -0,0 +1,35 @@
S-Lab License 1.0
Copyright 2022 S-Lab
Redistribution and use for non-commercial purpose in source and
binary forms, with or without modification, are permitted provided
that the following conditions are met:
1. Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
2. Redistributions in binary form must reproduce the above copyright
notice, this list of conditions and the following disclaimer in
the documentation and/or other materials provided with the
distribution.
3. Neither the name of the copyright holder nor the names of its
contributors may be used to endorse or promote products derived
from this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
In the event that redistribution and/or use for commercial purpose in
source or binary forms, with or without modification is required,
please contact the contributor(s) of the work.

437
LICENSE2 Normal file
View File

@ -0,0 +1,437 @@
Attribution-NonCommercial-ShareAlike 4.0 International
=======================================================================
Creative Commons Corporation ("Creative Commons") is not a law firm and
does not provide legal services or legal advice. Distribution of
Creative Commons public licenses does not create a lawyer-client or
other relationship. Creative Commons makes its licenses and related
information available on an "as-is" basis. Creative Commons gives no
warranties regarding its licenses, any material licensed under their
terms and conditions, or any related information. Creative Commons
disclaims all liability for damages resulting from their use to the
fullest extent possible.
Using Creative Commons Public Licenses
Creative Commons public licenses provide a standard set of terms and
conditions that creators and other rights holders may use to share
original works of authorship and other material subject to copyright
and certain other rights specified in the public license below. The
following considerations are for informational purposes only, are not
exhaustive, and do not form part of our licenses.
Considerations for licensors: Our public licenses are
intended for use by those authorized to give the public
permission to use material in ways otherwise restricted by
copyright and certain other rights. Our licenses are
irrevocable. Licensors should read and understand the terms
and conditions of the license they choose before applying it.
Licensors should also secure all rights necessary before
applying our licenses so that the public can reuse the
material as expected. Licensors should clearly mark any
material not subject to the license. This includes other CC-
licensed material, or material used under an exception or
limitation to copyright. More considerations for licensors:
wiki.creativecommons.org/Considerations_for_licensors
Considerations for the public: By using one of our public
licenses, a licensor grants the public permission to use the
licensed material under specified terms and conditions. If
the licensor's permission is not necessary for any reason--for
example, because of any applicable exception or limitation to
copyright--then that use is not regulated by the license. Our
licenses grant only permissions under copyright and certain
other rights that a licensor has authority to grant. Use of
the licensed material may still be restricted for other
reasons, including because others have copyright or other
rights in the material. A licensor may make special requests,
such as asking that all changes be marked or described.
Although not required by our licenses, you are encouraged to
respect those requests where reasonable. More considerations
for the public:
wiki.creativecommons.org/Considerations_for_licensees
=======================================================================
Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International
Public License
By exercising the Licensed Rights (defined below), You accept and agree
to be bound by the terms and conditions of this Creative Commons
Attribution-NonCommercial-ShareAlike 4.0 International Public License
("Public License"). To the extent this Public License may be
interpreted as a contract, You are granted the Licensed Rights in
consideration of Your acceptance of these terms and conditions, and the
Licensor grants You such rights in consideration of benefits the
Licensor receives from making the Licensed Material available under
these terms and conditions.
Section 1 -- Definitions.
a. Adapted Material means material subject to Copyright and Similar
Rights that is derived from or based upon the Licensed Material
and in which the Licensed Material is translated, altered,
arranged, transformed, or otherwise modified in a manner requiring
permission under the Copyright and Similar Rights held by the
Licensor. For purposes of this Public License, where the Licensed
Material is a musical work, performance, or sound recording,
Adapted Material is always produced where the Licensed Material is
synched in timed relation with a moving image.
b. Adapter's License means the license You apply to Your Copyright
and Similar Rights in Your contributions to Adapted Material in
accordance with the terms and conditions of this Public License.
c. BY-NC-SA Compatible License means a license listed at
creativecommons.org/compatiblelicenses, approved by Creative
Commons as essentially the equivalent of this Public License.
d. Copyright and Similar Rights means copyright and/or similar rights
closely related to copyright including, without limitation,
performance, broadcast, sound recording, and Sui Generis Database
Rights, without regard to how the rights are labeled or
categorized. For purposes of this Public License, the rights
specified in Section 2(b)(1)-(2) are not Copyright and Similar
Rights.
e. Effective Technological Measures means those measures that, in the
absence of proper authority, may not be circumvented under laws
fulfilling obligations under Article 11 of the WIPO Copyright
Treaty adopted on December 20, 1996, and/or similar international
agreements.
f. Exceptions and Limitations means fair use, fair dealing, and/or
any other exception or limitation to Copyright and Similar Rights
that applies to Your use of the Licensed Material.
g. License Elements means the license attributes listed in the name
of a Creative Commons Public License. The License Elements of this
Public License are Attribution, NonCommercial, and ShareAlike.
h. Licensed Material means the artistic or literary work, database,
or other material to which the Licensor applied this Public
License.
i. Licensed Rights means the rights granted to You subject to the
terms and conditions of this Public License, which are limited to
all Copyright and Similar Rights that apply to Your use of the
Licensed Material and that the Licensor has authority to license.
j. Licensor means the individual(s) or entity(ies) granting rights
under this Public License.
k. NonCommercial means not primarily intended for or directed towards
commercial advantage or monetary compensation. For purposes of
this Public License, the exchange of the Licensed Material for
other material subject to Copyright and Similar Rights by digital
file-sharing or similar means is NonCommercial provided there is
no payment of monetary compensation in connection with the
exchange.
l. Share means to provide material to the public by any means or
process that requires permission under the Licensed Rights, such
as reproduction, public display, public performance, distribution,
dissemination, communication, or importation, and to make material
available to the public including in ways that members of the
public may access the material from a place and at a time
individually chosen by them.
m. Sui Generis Database Rights means rights other than copyright
resulting from Directive 96/9/EC of the European Parliament and of
the Council of 11 March 1996 on the legal protection of databases,
as amended and/or succeeded, as well as other essentially
equivalent rights anywhere in the world.
n. You means the individual or entity exercising the Licensed Rights
under this Public License. Your has a corresponding meaning.
Section 2 -- Scope.
a. License grant.
1. Subject to the terms and conditions of this Public License,
the Licensor hereby grants You a worldwide, royalty-free,
non-sublicensable, non-exclusive, irrevocable license to
exercise the Licensed Rights in the Licensed Material to:
a. reproduce and Share the Licensed Material, in whole or
in part, for NonCommercial purposes only; and
b. produce, reproduce, and Share Adapted Material for
NonCommercial purposes only.
2. Exceptions and Limitations. For the avoidance of doubt, where
Exceptions and Limitations apply to Your use, this Public
License does not apply, and You do not need to comply with
its terms and conditions.
3. Term. The term of this Public License is specified in Section
6(a).
4. Media and formats; technical modifications allowed. The
Licensor authorizes You to exercise the Licensed Rights in
all media and formats whether now known or hereafter created,
and to make technical modifications necessary to do so. The
Licensor waives and/or agrees not to assert any right or
authority to forbid You from making technical modifications
necessary to exercise the Licensed Rights, including
technical modifications necessary to circumvent Effective
Technological Measures. For purposes of this Public License,
simply making modifications authorized by this Section 2(a)
(4) never produces Adapted Material.
5. Downstream recipients.
a. Offer from the Licensor -- Licensed Material. Every
recipient of the Licensed Material automatically
receives an offer from the Licensor to exercise the
Licensed Rights under the terms and conditions of this
Public License.
b. Additional offer from the Licensor -- Adapted Material.
Every recipient of Adapted Material from You
automatically receives an offer from the Licensor to
exercise the Licensed Rights in the Adapted Material
under the conditions of the Adapter's License You apply.
c. No downstream restrictions. You may not offer or impose
any additional or different terms or conditions on, or
apply any Effective Technological Measures to, the
Licensed Material if doing so restricts exercise of the
Licensed Rights by any recipient of the Licensed
Material.
6. No endorsement. Nothing in this Public License constitutes or
may be construed as permission to assert or imply that You
are, or that Your use of the Licensed Material is, connected
with, or sponsored, endorsed, or granted official status by,
the Licensor or others designated to receive attribution as
provided in Section 3(a)(1)(A)(i).
b. Other rights.
1. Moral rights, such as the right of integrity, are not
licensed under this Public License, nor are publicity,
privacy, and/or other similar personality rights; however, to
the extent possible, the Licensor waives and/or agrees not to
assert any such rights held by the Licensor to the limited
extent necessary to allow You to exercise the Licensed
Rights, but not otherwise.
2. Patent and trademark rights are not licensed under this
Public License.
3. To the extent possible, the Licensor waives any right to
collect royalties from You for the exercise of the Licensed
Rights, whether directly or through a collecting society
under any voluntary or waivable statutory or compulsory
licensing scheme. In all other cases the Licensor expressly
reserves any right to collect such royalties, including when
the Licensed Material is used other than for NonCommercial
purposes.
Section 3 -- License Conditions.
Your exercise of the Licensed Rights is expressly made subject to the
following conditions.
a. Attribution.
1. If You Share the Licensed Material (including in modified
form), You must:
a. retain the following if it is supplied by the Licensor
with the Licensed Material:
i. identification of the creator(s) of the Licensed
Material and any others designated to receive
attribution, in any reasonable manner requested by
the Licensor (including by pseudonym if
designated);
ii. a copyright notice;
iii. a notice that refers to this Public License;
iv. a notice that refers to the disclaimer of
warranties;
v. a URI or hyperlink to the Licensed Material to the
extent reasonably practicable;
b. indicate if You modified the Licensed Material and
retain an indication of any previous modifications; and
c. indicate the Licensed Material is licensed under this
Public License, and include the text of, or the URI or
hyperlink to, this Public License.
2. You may satisfy the conditions in Section 3(a)(1) in any
reasonable manner based on the medium, means, and context in
which You Share the Licensed Material. For example, it may be
reasonable to satisfy the conditions by providing a URI or
hyperlink to a resource that includes the required
information.
3. If requested by the Licensor, You must remove any of the
information required by Section 3(a)(1)(A) to the extent
reasonably practicable.
b. ShareAlike.
In addition to the conditions in Section 3(a), if You Share
Adapted Material You produce, the following conditions also apply.
1. The Adapter's License You apply must be a Creative Commons
license with the same License Elements, this version or
later, or a BY-NC-SA Compatible License.
2. You must include the text of, or the URI or hyperlink to, the
Adapter's License You apply. You may satisfy this condition
in any reasonable manner based on the medium, means, and
context in which You Share Adapted Material.
3. You may not offer or impose any additional or different terms
or conditions on, or apply any Effective Technological
Measures to, Adapted Material that restrict exercise of the
rights granted under the Adapter's License You apply.
Section 4 -- Sui Generis Database Rights.
Where the Licensed Rights include Sui Generis Database Rights that
apply to Your use of the Licensed Material:
a. for the avoidance of doubt, Section 2(a)(1) grants You the right
to extract, reuse, reproduce, and Share all or a substantial
portion of the contents of the database for NonCommercial purposes
only;
b. if You include all or a substantial portion of the database
contents in a database in which You have Sui Generis Database
Rights, then the database in which You have Sui Generis Database
Rights (but not its individual contents) is Adapted Material,
including for purposes of Section 3(b); and
c. You must comply with the conditions in Section 3(a) if You Share
all or a substantial portion of the contents of the database.
For the avoidance of doubt, this Section 4 supplements and does not
replace Your obligations under this Public License where the Licensed
Rights include other Copyright and Similar Rights.
Section 5 -- Disclaimer of Warranties and Limitation of Liability.
a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE
EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS
AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF
ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,
IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,
WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR
PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,
ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT
KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT
ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.
b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE
TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,
NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,
INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,
COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR
USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN
ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR
DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR
IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.
c. The disclaimer of warranties and limitation of liability provided
above shall be interpreted in a manner that, to the extent
possible, most closely approximates an absolute disclaimer and
waiver of all liability.
Section 6 -- Term and Termination.
a. This Public License applies for the term of the Copyright and
Similar Rights licensed here. However, if You fail to comply with
this Public License, then Your rights under this Public License
terminate automatically.
b. Where Your right to use the Licensed Material has terminated under
Section 6(a), it reinstates:
1. automatically as of the date the violation is cured, provided
it is cured within 30 days of Your discovery of the
violation; or
2. upon express reinstatement by the Licensor.
For the avoidance of doubt, this Section 6(b) does not affect any
right the Licensor may have to seek remedies for Your violations
of this Public License.
c. For the avoidance of doubt, the Licensor may also offer the
Licensed Material under separate terms or conditions or stop
distributing the Licensed Material at any time; however, doing so
will not terminate this Public License.
d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
License.
Section 7 -- Other Terms and Conditions.
a. The Licensor shall not be bound by any additional or different
terms or conditions communicated by You unless expressly agreed.
b. Any arrangements, understandings, or agreements regarding the
Licensed Material not stated herein are separate from and
independent of the terms and conditions of this Public License.
Section 8 -- Interpretation.
a. For the avoidance of doubt, this Public License does not, and
shall not be interpreted to, reduce, limit, restrict, or impose
conditions on any use of the Licensed Material that could lawfully
be made without permission under this Public License.
b. To the extent possible, if any provision of this Public License is
deemed unenforceable, it shall be automatically reformed to the
minimum extent necessary to make it enforceable. If the provision
cannot be reformed, it shall be severed from this Public License
without affecting the enforceability of the remaining terms and
conditions.
c. No term or condition of this Public License will be waived and no
failure to comply consented to unless expressly agreed to by the
Licensor.
d. Nothing in this Public License constitutes or may be interpreted
as a limitation upon, or waiver of, any privileges and immunities
that apply to the Licensor or You, including from the legal
processes of any jurisdiction or authority.
=======================================================================
Creative Commons is not a party to its public
licenses. Notwithstanding, Creative Commons may elect to apply one of
its public licenses to material it publishes and in those instances
will be considered the “Licensor.” The text of the Creative Commons
public licenses is dedicated to the public domain under the CC0 Public
Domain Dedication. Except for the limited purpose of indicating that
material is shared under a Creative Commons public license or as
otherwise permitted by the Creative Commons policies published at
creativecommons.org/policies, Creative Commons does not authorize the
use of the trademark "Creative Commons" or any other trademark or logo
of Creative Commons without its prior written consent including,
without limitation, in connection with any unauthorized modifications
to any of its public licenses or any other arrangements,
understandings, or agreements concerning use of licensed material. For
the avoidance of doubt, this paragraph does not form part of the
public licenses.
Creative Commons may be contacted at creativecommons.org.

122
README.md Normal file
View File

@ -0,0 +1,122 @@
# StableSR for Stable Diffusion WebUI
Licensed under S-Lab License 1.0
[![CC BY-NC-SA 4.0][cc-by-nc-sa-shield]][cc-by-nc-sa]
English｜[中文](README_CN.md)
- StableSR is a competitive super-resolution method originally proposed by Jianyi Wang et al.
- This repository is a migration of the StableSR project to the Automatic1111 WebUI.
Relevant Links
> Click to view high-quality official examples!
- [Project Page](https://iceclear.github.io/projects/stablesr/)
- [Official Repository](https://github.com/IceClear/StableSR)
- [Paper on arXiv](https://arxiv.org/abs/2305.07015)
> If you find this project useful, please give me & Jianyi Wang a star! ⭐
---
## Usage
### 1. Installation
⚪ Method 1: URL Install
- Open Automatic1111 WebUI -> Click Tab "Extensions" -> Click Tab "Install from URL" -> type in https://github.com/pkuliyi2015/sd-webui-stablesr.git -> Click "Install"
![installation](https://github.com/pkuliyi2015/multidiffusion-img-demo/blob/master/installation.png?raw=true)
⚪ Method 2: In progress...
> After sucessful installation, you should see "StableSR" in img2img Scripts dropdown list.
### 2. Download the main components
- You MUST use the Stable Diffusion V2.1 512 **EMA** checkpoint (~5.21GB) from StabilityAI
- You can download it from [HuggingFace](https://huggingface.co/stabilityai/stable-diffusion-2-1-base)
- Put into stable-diffusion-webui/models/Stable-Diffusion/
- Download the pruned StableSR module (~
400MB)
- Official resources: In Progress
- My resources: <[GoogleDrive](https://drive.google.com/file/d/1tWjkZQhfj07sHDR4r9Ta5Fk4iMp1t3Qw/view?usp=sharing)> <[百度网盘-提取码aguq](https://pan.baidu.com/s/1Nq_6ciGgKnTu0W14QcKKWg?pwd=aguq)>
- Put into stable-diffusion-webui/extensions/sd-webui-stablesr/models/
### 3. Optional components
- Install [Tiled Diffusion & VAE]((https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111)) extension
- The original StableSR easily gets OOM for large images > 512.
- For better quality and less VRAM usage, we recommend Tiled Diffusion & VAE.
- Use the Official VQGAN VAE (~700MB)
- Official resources: In Progress
- My resources: <[GoogleDrive](https://drive.google.com/file/d/1ARtDMia3_CbwNsGxxGcZ5UP75W4PeIEI/view?usp=share_link)> <[百度网盘-提取码83u9](https://pan.baidu.com/s/1YCYmGBethR9JZ8-eypoIiQ?pwd=83u9)>
- Put it in your stable-diffusion-webui/models/VAE
### 4. Extension Usage
- At the top of the WebUI, select the v2-1_512-ema-pruned checkpoint you downloaded.
- Switch to img2img tag. Find the "Scripts" dropdown at the bottom of the page.
- Select the StableSR script.
- Click the refresh button and select the StableSR checkpoint you have downloaded.
- Choose a scale factor.
- Upload your image and start generation (can work without prompts).
### 5. Useful Tips
- Euler a sampler is recommended. Steps >= 20.
- For output image size > 512, we recommend using Tiled Diffusion & VAE, otherwise, the image quality may not be ideal, and the VRAM usage will be huge.
- Here are the Tiled Diffusion settings that replicate the official behavior in the paper.
- Method = Mixture of Diffusers
- Latent tile size = 64, Latent tile overlap = 32
- Latent tile batch size as large as possible before Out of Memory.
- Upscaler MUST be None.
- What is "Pure Noise"?
- Pure Noise refers to starting from a fully random noise tensor instead of your image. **This is the default behavior in the StableSR paper.**
- When enabling it, the script ignores your denoising strength and gives you much more detailed images, but also changes the color & sharpness significantly
- When disabling it, the script starts by adding some noise to your image. The result will be not fully detailed, even if you set denoising strength = 1 (but maybe aesthetically good). See [Comparison](https://imgsli.com/MTgwMTMx).
### 6. Important Notice
> Why my results are different from the offical examples?
- It is not your or our fault.
- This extension has the same UNet model weights as the StableSR if installed correctly.
- If you install the optional VQVAE, the whole model weights will be the same as the official model with fusion weights=0.
- However, your result will be **not as good as** the official results, because:
- Sampler Difference:
- The official repo does 100 or 200 steps of legacy DDPM sampling with a custom timestep scheduler, and samples without negative prompts.
- However, WebUI doesn't offer such a sampler, and it must sample with negative prompts. **This is the main difference.**
- VQVAE Decoder Difference:
- The official VQVAE Decoder takes some Encoder features as input.
- However, in practice, I found these features are astonishingly huge for large images. (>10G for 4k images even in float16!)
- Hence, **I removed the CFW component in VAE Decoder**. As this lead to inferior fidelity in details, I will try to add it back later as an option.
---
## License
This project is licensed under:
- S-Lab License 1.0.
- [Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License][cc-by-nc-sa], due to the use of the NVIDIA SPADE module.
[![CC BY-NC-SA 4.0][cc-by-nc-sa-image]][cc-by-nc-sa]
[cc-by-nc-sa]: http://creativecommons.org/licenses/by-nc-sa/4.0/
[cc-by-nc-sa-image]: https://licensebuttons.net/l/by-nc-sa/4.0/88x31.png
[cc-by-nc-sa-shield]: https://img.shields.io/badge/License-CC%20BY--NC--SA%204.0-lightgrey.svg
### Disclaimer
- All code in this extension is for research purposes only.
- The commercial use of the code and checkpoint is **strictly prohibited**.
### Important Notice for Outcome Images
- Please note that the CC BY-NC-SA 4.0 license in the NVIDIA SPADE module also prohibits the commercial use of outcome images.
- Jianyi Wang may change the SPADE module to a commercial-friendly one but he is busy.
- If you wish to *speed up* his process for commercial purposes, please contact him through email: iceclearwjy@gmail.com
## Acknowledgments
I would like to thank Jianyi Wang et al. for the original StableSR method.

119
README_CN.md Normal file
View File

@ -0,0 +1,119 @@
# StableSR - Stable Diffusion WebUI
S-Lab License 1.0 & [![CC BY-NC-SA 4.0][cc-by-nc-sa-shield]][cc-by-nc-sa]
[English](README.md) | 中文
- StableSR 是原初由 Jianyi Wang 等人提出的具有竞争力的超分辨率方法。
- 本仓库是将 StableSR 项目迁移到 Automatic1111 WebUI 的迁移工作。
相关链接
> 点击查看高质量官方示例！
- [项目页面](https://iceclear.github.io/projects/stablesr/)
- [官方仓库](https://github.com/IceClear/StableSR)
- [arXiv 上的论文](https://arxiv.org/abs/2305.07015)
> 如果你觉得这个项目有用，请给我和 Jianyi Wang 点个赞！⭐
---
## 使用
### 1. 安装
⚪ 方法 1: URL 安装
- 打开 Automatic1111 WebUI -> 点击 "扩展" 标签页 -> 点击 "从 URL 安装" 标签页 -> 输入 https://github.com/pkuliyi2015/sd-webui-stablesr.git -> 点击 "安装"
![installation](https://github.com/pkuliyi2015/multidiffusion-img-demo/blob/master/installation.png?raw=true)
⚪ 方法 2: 进行中...
> 安装成功后，你应该能在 img2img 脚本下拉列表中看到 "StableSR"。
### 2. 下载主要组件
- 你必须使用来自 StabilityAI 的 Stable Diffusion V2.1 512 **EMA** 检查点（大约 5.21GB）
- 你可以从 [HuggingFace](https://huggingface.co/stabilityai/stable-diffusion-2-1-base) 下载它
- 放入 stable-diffusion-webui/models/Stable-Diffusion/
- 下载剪枝后的 StableSR 模块（大约 400MB）
- 官方资源：进行中
- 我的资源：<[GoogleDrive](https://drive.google.com/file/d/1tWjkZQhfj07sHDR4r9Ta5Fk4iMp1t3Qw/view?usp=sharing)> <[百度网盘-提取码aguq](https://pan.baidu.com/s/1Nq_6ciGgKnTu0W14QcKKWg?pwd=aguq)>
- 放入 stable-diffusion-webui/extensions/sd-webui-stablesr/models/
### 3. 可选组件
- 安装 [Tiled Diffusion & VAE](https://github.com/pkuliyi2015/multidiffusion-upscaler-for-automatic1111) 扩展
- 原始的 StableSR 对大于 512 的大图像容易出现 OOM。
- 为了获得更好的质量和更少的 VRAM 使用，我们建议使用 Tiled Diffusion & VAE。
- 使用官方 VQGAN VAE（大约 700MB）
- 官方资源：进行中
- 我的资源：<[GoogleDrive](https://drive.google.com/file/d/1ARtDMia3_CbwNsGxxGcZ5UP75W4PeIEI/view?usp=share_link)> <[百度网盘-提取码83u9](https://pan.baidu.com/s/1YCYmGBethR9JZ8-eypoIiQ?pwd=83u9)>
- 将它放在你的 stable-diffusion-webui/models/VAE 中
### 4. 扩展使用
- 在 WebUI 的顶部，选择你下载的 v2-1_512-ema-pruned 检查点。
- 切换到 img2img 标签。在页面底部找到 "脚本" 下拉列表。
- 选择 StableSR 脚本。
- 点击刷新按钮并选择你已下载的 StableSR 检查点。
- 选择一个比例因子。
- 上传你的图像并开始生成（无需提示）。
### 5. 有用的提示
- 推荐使用 Euler 采样器。步数 >= 20。
- 对于输出图像大小 > 512，我们推荐使用 Tiled Diffusion & VAE，否则，图像质量可能不理想，VRAM 使用量会很大。
- 这里有一些 Tiled Diffusion 设置，可以复制论文中的官方行为。
- 方法 = Diffusers 混合
- 隐变量瓷砖大小 = 64，隐变量瓷砖重叠 = 32
- 隐变量瓷砖批大小尽可能大，避免内存不足。
- 上采样器必须为 None。
- 什么是 "纯噪声"？
- 纯噪声指的是从完全随机的噪声张量开始，而不是从你的图像开始。**这是 StableSR 论文中的默认行为。**
- 启用时，脚本会忽略你的去噪强度，并给你更详细的图像，但也会显著改变颜色和锐度
- 禁用时，脚本会开始添加一些噪声到你的图像。即使你将去噪强度设为 1，结果也不会完全详细（但可能在美感上更好）。参见 [对比](https://imgsli.com/MTgwMTMx)。
### 6. 重要提醒
> 为什么我的结果和官方示例不同？
- 这不是你或我们的错。
- 如果正确安装，这个扩展有与 StableSR 相同的 UNet 模型权重。
- 如果你安装了可选的 VQVAE，整个模型权重将与融合权重为 0 的官方模型相同。
- 但是，你的结果将**不如**官方结果，因为：
- 采样器差异：
-官方仓库进行 100 或 200 步的 legacy DDPM 采样，并使用自定义的时间步调度器，采样时不使用负提示。
- 然而，WebUI 不提供这样的采样器，必须带有负提示进行采样。**这是主要的差异。**
- VQVAE 解码器差异：
- 官方 VQVAE 解码器将一些编码器特征作为输入。
- 然而，在实践中，我发现这些特征对于大图像来说非常大。 (>10G 用于 4k 图像，即使是在 float16！)
- 因此，**我移除了 VAE 解码器中的 CFW 组件**。由于这导致了对细节的较低保真度，我将尝试将它作为一个选项添加回去。
---
## 许可
此项目在以下许可下授权：
- S-Lab License 1.0.
- [Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License][cc-by-nc-sa]，由于使用了 NVIDIA SPADE 模块。
[![CC BY-NC-SA 4.0][cc-by-nc-sa-image]][cc-by-nc-sa]
[cc-by-nc-sa]: http://creativecommons.org/licenses/by-nc-sa/4.0/
[cc-by-nc-sa-image]: https://licensebuttons.net/l/by-nc-sa/4.0/88x31.png
[cc-by-nc-sa-shield]: https://img.shields.io/badge/License-CC%20BY--NC--SA%204.0-lightgrey.svg
### 免责声明
- 此扩展中的所有代码仅供研究目的。
- 代码和检查点的商业用途**严格禁止**。
### 成果图像的重要通知
- 请注意，NVIDIA SPADE 模块中的 CC BY-NC-SA 4.0 许可也禁止使用成果图像进行商业用途。
- Jianyi Wang 可能会将 SPADE 模块更改为商业友好的一个，但他很忙。
- 如果你希望*加快*他为商业目的的进程，请通过电子邮件与他联系：iceclearwjy@gmail.com
## 致谢
我要感谢 Jianyi Wang 等人提出的原始 StableSR 方法。

238
scripts/stablesr.py Normal file
View File

@ -0,0 +1,238 @@
'''
# --------------------------------------------------------------------------------
#
# StableSR for Automatic1111 WebUI
#
# Introducing state-of-the super-resolution method: StableSR!
# Techniques is originally proposed by my schoolmate Jianyi Wang et, al.
#
# Project Page: https://iceclear.github.io/projects/stablesr/
# Official Repo: https://github.com/IceClear/StableSR
# Paper: https://arxiv.org/abs/2305.07015
#
# @original author: Jianyi Wang et, al.
# @migration: LI YI
# @organization: Nanyang Technological University - Singapore
# @date: 2023-05-20
# @license:
# S-Lab License 1.0 (see LICENSE file)
# CC BY-NC-SA 4.0 (required by NVIDIA SPADE module)
#
# @disclaimer:
# All code in this extension is for research purpose only.
# The commercial use of the code & checkpoint is strictly prohibited.
#
# --------------------------------------------------------------------------------
#
# IMPORTANT NOTICE FOR OUTCOME IMAGES:
# - Please be aware that the CC BY-NC-SA 4.0 license in SPADE module
# also prohibits the commercial use of outcome images.
# - Jianyi Wang may change the SPADE module to a commercial-friendly one.
# If you want to use the outcome images for commercial purposes, please
# contact Jianyi Wang for more information.
#
# Please give me a star (and also Jianyi's repo) if you like this project!
#
# --------------------------------------------------------------------------------
'''
import os
import torch
import gradio as gr
import numpy as np
import PIL.Image as Image
from pathlib import Path
from torch import Tensor
from tqdm import tqdm
from modules import scripts, processing, sd_samplers, devices
from modules.processing import StableDiffusionProcessingImg2Img, Processed
from ldm.modules.diffusionmodules.openaimodel import UNetModel
from srmodule.spade import SPADELayers
from srmodule.struct_cond import EncoderUNetModelWT, build_unetwt
from srmodule.colorfix import fix_color
SD_WEBUI_PATH = Path.cwd()
ME_PATH = SD_WEBUI_PATH / 'extensions' / 'sd-webui-stablesr'
MODEL_PATH = ME_PATH / 'models'
FORWARD_CACHE_NAME = 'org_forward_stablesr'
class StableSR:
def __init__(self, path, dtype, device):
state_dict = torch.load(path, map_location='cpu')
self.struct_cond_model: EncoderUNetModelWT = build_unetwt()
self.spade_layers: SPADELayers = SPADELayers()
self.struct_cond_model.load_from_dict(state_dict)
self.spade_layers.load_from_dict(state_dict)
del state_dict
self.struct_cond_model.apply(lambda x: x.to(dtype=dtype, device=device))
self.spade_layers.apply(lambda x: x.to(dtype=dtype, device=device))
self.latent_image: Tensor = None
self.set_image_hooks = {}
self.struct_cond: Tensor = None
def set_latent_image(self, latent_image):
self.latent_image = latent_image
for hook in self.set_image_hooks.values():
hook(latent_image)
def hook(self, unet: UNetModel):
# hook unet to set the struct_cond
if not hasattr(unet, FORWARD_CACHE_NAME):
setattr(unet, FORWARD_CACHE_NAME, unet.forward)
def unet_forward(x, timesteps=None, context=None, y=None,**kwargs):
self.latent_image = self.latent_image.to(x.device)
self.struct_cond = None # mitigate vram peak
self.struct_cond = self.struct_cond_model(self.latent_image, timesteps.to(x.device)[:self.latent_image.shape[0]])
return getattr(unet, FORWARD_CACHE_NAME)(x, timesteps, context, y, **kwargs)
unet.forward = unet_forward
self.spade_layers.hook(unet, lambda: self.struct_cond)
def unhook(self, unet: UNetModel):
# clean up cache
self.latent_image = None
self.struct_cond = None
self.set_image_hooks = {}
# unhook unet forward
if hasattr(unet, FORWARD_CACHE_NAME):
unet.forward = getattr(unet, FORWARD_CACHE_NAME)
delattr(unet, FORWARD_CACHE_NAME)
# unhook spade layers
self.spade_layers.unhook(unet)
class Script(scripts.Script):
def __init__(self) -> None:
self.model_list = {}
self.load_model_list()
self.last_path = None
self.stablesr_model: StableSR = None
def load_model_list(self):
# traverse the CFG_PATH and add all files to the model list
self.model_list = {}
for file in MODEL_PATH.iterdir():
if file.is_file():
# save tha absolute path
self.model_list[file.name] = str(file.absolute())
self.model_list['None'] = None
def title(self):
return "StableSR"
def show(self, is_img2img):
return is_img2img
def ui(self, is_img2img):
gr.HTML('<p>StableSR is a state-of-the-art super-resolution method.</p>')
gr.HTML('<p>1. You MUST use SD2.1-512-ema-pruned checkpoint. Euler a sampler is recommended.</p>')
gr.HTML('<p>2. Use Tiled Diffusion & VAE - Mixture of Diffusers for resolution > 512.</p>')
gr.HTML('<p>3. When use Tiled Diffusion, you MUST set the upscaler to None!</p>')
with gr.Row():
model = gr.Dropdown(list(self.model_list.keys()), label="SR Model")
refresh = gr.Button(value='↻', variant='tool')
def refresh_fn(selected):
self.load_model_list()
if selected not in self.model_list:
selected = 'None'
return gr.Dropdown.update(value=selected, choices=list(self.model_list.keys()))
refresh.click(fn=refresh_fn,inputs=model, outputs=model)
with gr.Row():
scale_factor = gr.Slider(minimum=1, maximum=16, step=0.1, value=2, label='Scale Factor', elem_id=f'StableSR-scale')
with gr.Row():
pure_noise = gr.Checkbox(label='Pure Noise', value=True, elem_id=f'StableSR-pure-noise')
color_fix = gr.Checkbox(label='Color Fix', value=True, elem_id=f'StableSR-color-fix')
return [model, scale_factor, pure_noise, color_fix]
def run(self, p: StableDiffusionProcessingImg2Img, model: str, scale_factor:float, pure_noise: bool, color_fix:bool):
if model == 'None':
# do clean up
self.stablesr_model = None
self.last_model_path = None
return
if model not in self.model_list:
raise gr.Error(f"Model {model} is not in the list! Please refresh your browser!")
if not os.path.exists(self.model_list[model]):
raise gr.Error(f"Model {model} is not on your disk! Please refresh the model list!")
# upscale the image, set the ouput size
init_img: Image = p.init_images[0]
target_width = int(init_img.width * scale_factor)
target_height = int(init_img.height * scale_factor)
# if the target width is not dividable by 8, then round it up
if target_width % 8 != 0:
target_width = target_width + 8 - target_width % 8
# if the target height is not dividable by 8, then round it up
if target_height % 8 != 0:
target_height = target_height + 8 - target_height % 8
init_img = init_img.resize((target_width, target_height), Image.LANCZOS)
p.init_images[0] = init_img
p.width = init_img.width
p.height = init_img.height
print('[StableSR] Target image size: {}x{}'.format(init_img.width, init_img.height))
unet: UNetModel = p.sd_model.model.diffusion_model
# print(unet.input_blocks)
first_param = unet.parameters().__next__()
if self.last_path != self.model_list[model]:
# load the model
self.stablesr_model = None
# get the type and the device of the unet model's first parameter
self.stablesr_model = StableSR(self.model_list[model], dtype=first_param.dtype, device=first_param.device)
self.last_path = self.model_list[model]
def sample_custom(conditioning, unconditional_conditioning, seeds, subseeds, subseed_strength, prompts):
self.stablesr_model.set_latent_image(p.init_latent)
x = processing.create_random_tensors(p.init_latent.shape[1:], seeds=seeds, subseeds=subseeds, subseed_strength=p.subseed_strength, seed_resize_from_h=p.seed_resize_from_h, seed_resize_from_w=p.seed_resize_from_w, p=p)
sampler = sd_samplers.create_sampler(p.sampler_name, p.sd_model)
if pure_noise:
# NOTE: use txt2img instead of img2img sampling
samples = sampler.sample(p, x, conditioning, unconditional_conditioning, image_conditioning=p.image_conditioning)
else:
if p.initial_noise_multiplier != 1.0:
p.extra_generation_params["Noise multiplier"] =p.initial_noise_multiplier
x *= p.initial_noise_multiplier
samples = sampler.sample_img2img(p, p.init_latent, x, conditioning, unconditional_conditioning, image_conditioning=p.image_conditioning)
if p.mask is not None:
samples = samples * p.nmask + p.init_latent * p.mask
del x
devices.torch_gc()
return samples
# replace the sample function
p.sample = sample_custom
# Hook the unet, and unhook after processing.
try:
self.stablesr_model.hook(unet)
result: Processed = processing.process_images(p)
if color_fix:
for i in range(len(result.images)):
result.images[i] = fix_color(result.images[i], init_img)
return result
finally:
self.stablesr_model.unhook(unet)

111
srmodule/attn.py Normal file
View File

@ -0,0 +1,111 @@
'''
This file is modified from the TiledVAE attn.py, so that the StableSR can save much VRAM.
'''
import math
import torch
from modules import shared, sd_hijack
from modules.sd_hijack_optimizations import get_available_vram, get_xformers_flash_attention_op, sub_quad_attention
try:
import xformers
import xformers.ops
except ImportError:
pass
def get_attn_func():
method = sd_hijack.model_hijack.optimization_method
if method is None:
return attn_forward
method = method.lower()
# The method should be one of the following:
# ['none', 'sdp-no-mem', 'sdp', 'xformers', ''sub-quadratic', 'v1', 'invokeai', 'doggettx']
if method not in ['none', 'sdp-no-mem', 'sdp', 'xformers', 'sub-quadratic', 'v1', 'invokeai', 'doggettx']:
print(f"[StableSR] Warning: Unknown attention optimization method {method}. Please try to update the extension.")
return attn_forward
if method == 'none':
return attn_forward
elif method == 'xformers':
return xformers_attnblock_forward
elif method == 'sdp-no-mem':
return sdp_no_mem_attnblock_forward
elif method == 'sdp':
return sdp_attnblock_forward
elif method == 'sub-quadratic':
return sub_quad_attnblock_forward
elif method == 'doggettx':
return cross_attention_attnblock_forward
return attn_forward
# The following functions are all copied from modules.sd_hijack_optimizations
# However, the residual & normalization are removed and computed separately.
def attn_forward(q, k, v):
# compute attention
# q: b,hw,c
k = k.permute(0, 2, 1) # b,c,hw
c = k.shape[1]
w_ = torch.bmm(q, k) # b,hw,hw w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
w_ = w_ * (int(c)**(-0.5))
w_ = torch.nn.functional.softmax(w_, dim=2)
# attend to values
v = v.permute(0, 2, 1) # b,c,hw
w_ = w_.permute(0, 2, 1) # b,hw,hw (first hw of k, second of q)
# b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j]
h_ = torch.bmm(v, w_)
return h_.permute(0, 2, 1)
def xformers_attnblock_forward(q, k, v):
return xformers.ops.memory_efficient_attention(q, k, v, op=get_xformers_flash_attention_op(q, k, v))
def cross_attention_attnblock_forward(q, k, v):
# compute attention
k = k.permute(0, 2, 1)# b,c,hw
v = v.permute(0, 2, 1)# b,c,hw
c = k.shape[1]
h_ = torch.zeros_like(k, device=q.device)
mem_free_total = get_available_vram()
tensor_size = q.shape[0] * q.shape[1] * k.shape[2] * q.element_size()
mem_required = tensor_size * 2.5
steps = 1
if mem_required > mem_free_total:
steps = 2**(math.ceil(math.log(mem_required / mem_free_total, 2)))
slice_size = q.shape[1] // steps if (q.shape[1] % steps) == 0 else q.shape[1]
for i in range(0, q.shape[1], slice_size):
end = i + slice_size
w1 = torch.bmm(q[:, i:end], k) # b,hw,hw w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
w2 = w1 * (int(c)**(-0.5))
del w1
w3 = torch.nn.functional.softmax(w2, dim=2, dtype=q.dtype)
del w2
# attend to values
w4 = w3.permute(0, 2, 1) # b,hw,hw (first hw of k, second of q)
del w3
h_[:, :, i:end] = torch.bmm(v, w4) # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j]
del w4
return h_.permute(0, 2, 1)
def sdp_no_mem_attnblock_forward(q, k, v):
with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=True, enable_mem_efficient=False):
return sdp_attnblock_forward(q, k, v)
def sdp_attnblock_forward(q, k, v):
return torch.nn.functional.scaled_dot_product_attention(q, k, v, dropout_p=0.0, is_causal=False)
def sub_quad_attnblock_forward(q, k, v):
return sub_quad_attention(q, k, v, q_chunk_size=shared.cmd_opts.sub_quad_q_chunk_size, kv_chunk_size=shared.cmd_opts.sub_quad_kv_chunk_size, chunk_threshold=shared.cmd_opts.sub_quad_chunk_threshold, use_checkpoint=True)

48
srmodule/colorfix.py Normal file
View File

@ -0,0 +1,48 @@
from PIL import Image
from torch import Tensor
from torchvision.transforms import ToTensor, ToPILImage
def fix_color(target: Image, source: Image):
# Convert images to tensors
to_tensor = ToTensor()
target_tensor = to_tensor(target).unsqueeze(0)
source_tensor = to_tensor(source).unsqueeze(0)
# Apply adaptive instance normalization
result_tensor = adaptive_instance_normalization(target_tensor, source_tensor)
# Convert tensor back to image
to_image = ToPILImage()
result_image = to_image(result_tensor.squeeze(0).clamp_(0.0, 1.0))
return result_image
def calc_mean_std(feat: Tensor, eps=1e-5):
"""Calculate mean and std for adaptive_instance_normalization.
Args:
feat (Tensor): 4D tensor.
eps (float): A small value added to the variance to avoid
divide-by-zero. Default: 1e-5.
"""
size = feat.size()
assert len(size) == 4, 'The input feature should be 4D tensor.'
b, c = size[:2]
feat_var = feat.view(b, c, -1).var(dim=2) + eps
feat_std = feat_var.sqrt().view(b, c, 1, 1)
feat_mean = feat.view(b, c, -1).mean(dim=2).view(b, c, 1, 1)
return feat_mean, feat_std
def adaptive_instance_normalization(content_feat:Tensor, style_feat:Tensor):
"""Adaptive instance normalization.
Adjust the reference features to have the similar color and illuminations
as those in the degradate features.
Args:
content_feat (Tensor): The reference feature.
style_feat (Tensor): The degradate features.
"""
size = content_feat.size()
style_mean, style_std = calc_mean_std(style_feat)
content_mean, content_std = calc_mean_std(content_feat)
normalized_feat = (content_feat - content_mean.expand(size)) / content_std.expand(size)
return normalized_feat * style_std.expand(size) + style_mean.expand(size)

201
srmodule/spade.py Normal file
View File

@ -0,0 +1,201 @@
"""
Copyright (C) 2019 NVIDIA Corporation. All rights reserved.
Licensed under the CC BY-NC-SA 4.0 license (https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode).
"""
import re
import torch
import torch.nn as nn
from ldm.modules.diffusionmodules.util import normalization, checkpoint
from ldm.modules.diffusionmodules.openaimodel import ResBlock, UNetModel
class SPADE(nn.Module):
def __init__(self, norm_nc, label_nc=256, config_text='spadeinstance3x3'):
super().__init__()
assert config_text.startswith('spade')
parsed = re.search('spade(\D+)(\d)x\d', config_text)
ks = int(parsed.group(2))
self.param_free_norm = normalization(norm_nc)
# The dimension of the intermediate embedding space. Yes, hardcoded.
nhidden = 128
pw = ks // 2
self.mlp_shared = nn.Sequential(
nn.Conv2d(label_nc, nhidden, kernel_size=ks, padding=pw),
nn.ReLU()
)
self.mlp_gamma = nn.Conv2d(nhidden, norm_nc, kernel_size=ks, padding=pw)
self.mlp_beta = nn.Conv2d(nhidden, norm_nc, kernel_size=ks, padding=pw)
def forward(self, x_dic, segmap_dic):
return checkpoint(
self._forward, (x_dic, segmap_dic), self.parameters(), True
)
def _forward(self, x_dic, segmap_dic):
segmap = segmap_dic[str(x_dic.size(-1))]
x = x_dic
# Part 1. generate parameter-free normalized activations
normalized = self.param_free_norm(x)
# Part 2. produce scaling and bias conditioned on semantic map
# segmap = F.interpolate(segmap, size=x.size()[2:], mode='nearest')
actv = self.mlp_shared(segmap)
repeat_factor = normalized.shape[0]//segmap.shape[0]
if repeat_factor > 1:
out = normalized
out *= (1 + self.mlp_gamma(actv).repeat_interleave(repeat_factor, dim=0))
out += self.mlp_beta(actv).repeat_interleave(repeat_factor, dim=0)
else:
out = normalized
out *= (1 + self.mlp_gamma(actv))
out += self.mlp_beta(actv)
return out
def dual_resblock_forward(self: ResBlock, x, emb, spade: SPADE, get_struct_cond):
if self.updown:
in_rest, in_conv = self.in_layers[:-1], self.in_layers[-1]
h = in_rest(x)
h = self.h_upd(h)
x = self.x_upd(x)
h = in_conv(h)
else:
h = self.in_layers(x)
emb_out = self.emb_layers(emb).type(h.dtype)
while len(emb_out.shape) < len(h.shape):
emb_out = emb_out[..., None]
if self.use_scale_shift_norm:
out_norm, out_rest = self.out_layers[0], self.out_layers[1:]
scale, shift = torch.chunk(emb_out, 2, dim=1)
h = out_norm(h) * (1 + scale) + shift
h = out_rest(h)
else:
h = h + emb_out
h = self.out_layers(h)
h = spade(h, get_struct_cond())
return self.skip_connection(x) + h
class SPADELayers(nn.Module):
def __init__(self):
'''
A container class for fast SPADE layer loading.
params inferred from the official checkpoint
'''
super().__init__()
self.input_blocks = nn.ModuleList([
nn.Identity(),
SPADE(320),
SPADE(320),
nn.Identity(),
SPADE(640),
SPADE(640),
nn.Identity(),
SPADE(1280),
SPADE(1280),
nn.Identity(),
SPADE(1280),
SPADE(1280),
])
self.middle_block = nn.ModuleList([
SPADE(1280),
nn.Identity(),
SPADE(1280),
])
self.output_blocks = nn.ModuleList([
SPADE(1280),
SPADE(1280),
SPADE(1280),
SPADE(1280),
SPADE(1280),
SPADE(1280),
SPADE(640),
SPADE(640),
SPADE(640),
SPADE(320),
SPADE(320),
SPADE(320),
])
self.input_ids = [1,2,4,5,7,8,10,11]
self.output_ids = list(range(12))
self.mid_ids = [0,2]
self.forward_cache_name = 'org_forward_stablesr'
def hook(self, unet: UNetModel, get_struct_cond):
# hook all resblocks
resblock: ResBlock = None
for i in self.input_ids:
resblock = unet.input_blocks[i][0]
# debug
# assert isinstance(resblock, ResBlock)
if not hasattr(resblock, self.forward_cache_name):
setattr(resblock, self.forward_cache_name, resblock._forward)
resblock._forward = lambda x, timesteps, resblock=resblock, spade=self.input_blocks[i]: dual_resblock_forward(resblock, x, timesteps, spade, get_struct_cond)
for i in self.output_ids:
resblock = unet.output_blocks[i][0]
# debug
# assert isinstance(resblock, ResBlock)
if not hasattr(resblock, self.forward_cache_name):
setattr(resblock, self.forward_cache_name, resblock._forward)
resblock._forward = lambda x, timesteps, resblock=resblock, spade=self.output_blocks[i]: dual_resblock_forward(resblock, x, timesteps, spade, get_struct_cond)
for i in self.mid_ids:
resblock = unet.middle_block[i]
# debug
# assert isinstance(resblock, ResBlock)
if not hasattr(resblock, self.forward_cache_name):
setattr(resblock, self.forward_cache_name, resblock._forward)
resblock._forward = lambda x, timesteps, resblock=resblock, spade=self.middle_block[i]: dual_resblock_forward(resblock, x, timesteps, spade, get_struct_cond)
def unhook(self, unet: UNetModel):
resblock: ResBlock = None
for i in self.input_ids:
resblock = unet.input_blocks[i][0]
if hasattr(resblock, self.forward_cache_name):
resblock._forward = getattr(resblock, self.forward_cache_name)
delattr(resblock, self.forward_cache_name)
for i in self.output_ids:
resblock = unet.output_blocks[i][0]
if hasattr(resblock, self.forward_cache_name):
resblock._forward = getattr(resblock, self.forward_cache_name)
delattr(resblock, self.forward_cache_name)
for i in self.mid_ids:
resblock = unet.middle_block[i]
if hasattr(resblock, self.forward_cache_name):
resblock._forward = getattr(resblock, self.forward_cache_name)
delattr(resblock, self.forward_cache_name)
def load_from_dict(self, state_dict):
"""
Load model weights from a dictionary.
:param state_dict: a dict of parameters.
"""
filtered_dict = {}
for k, v in state_dict.items():
if k.startswith("model.diffusion_model."):
key = k[len("model.diffusion_model.") :]
# remove the '.0.spade' within the key
if 'middle_block' not in key:
key = key.replace('.0.spade', '')
else:
key = key.replace('.spade', '')
filtered_dict[key] = v
self.load_state_dict(filtered_dict)
if __name__ == '__main__':
path = '../models/stablesr_sd21.ckpt'
state_dict = torch.load(path)
model = SPADELayers()
model.load_from_dict(state_dict)
print(model)

353
srmodule/struct_cond.py Normal file
View File

@ -0,0 +1,353 @@
import math
import torch
import torch.nn as nn
from ldm.modules.diffusionmodules.openaimodel import (
TimestepEmbedSequential,
ResBlock,
Downsample,
)
from ldm.modules.diffusionmodules.util import (
conv_nd,
linear,
timestep_embedding,
checkpoint,
normalization,
zero_module,
)
from srmodule.attn import get_attn_func
attn_func = None
class QKVAttentionLegacy(nn.Module):
"""
A module which performs QKV attention. Matches legacy QKVAttention + input/ouput heads shaping
"""
def __init__(self, n_heads):
super().__init__()
self.n_heads = n_heads
def forward(self, qkv):
"""
Apply QKV attention.
:param qkv: an [N x (3 * H * C) x T] tensor of Qs, Ks, and Vs.
:return: an [N x (H * C) x T] tensor after attention.
"""
bs, width, length = qkv.shape
assert width % (3 * self.n_heads) == 0
ch = width // (3 * self.n_heads)
q, k, v = qkv.reshape(bs * self.n_heads, ch * 3, length).split(ch, dim=1)
# Legacy Attention
# scale = 1 / math.sqrt(math.sqrt(ch))
# weight = torch.einsum(
# "bct,bcs->bts", q * scale, k * scale
# ) # More stable with f16 than dividing afterwards
# weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
# a = torch.einsum("bts,bcs->bct", weight, v)
# a = a.reshape(bs, -1, length)
q, k, v = map(
lambda t:t.permute(0,2,1)
.contiguous(),
(q, k, v),
)
global attn_func
a = attn_func(q, k, v)
a = (
a.permute(0,2,1)
.reshape(bs, -1, length)
)
return a
class AttentionBlock(nn.Module):
"""
An attention block that allows spatial positions to attend to each other.
Originally ported from here, but adapted to the N-d case.
https://github.com/hojonathanho/diffusion/blob/1e0dceb3b3495bbe19116a5e1b3596cd0706c543/diffusion_tf/models/unet.py#L66.
"""
def __init__(
self,
channels,
num_heads=1,
num_head_channels=-1,
use_checkpoint=False,
use_new_attention_order=False,
):
super().__init__()
self.channels = channels
if num_head_channels == -1:
self.num_heads = num_heads
else:
assert (
channels % num_head_channels == 0
), f"q,k,v channels {channels} is not divisible by num_head_channels {num_head_channels}"
self.num_heads = channels // num_head_channels
self.norm = normalization(channels)
self.qkv = conv_nd(1, channels, channels * 3, 1)
self.attention = QKVAttentionLegacy(self.num_heads)
self.proj_out = zero_module(conv_nd(1, channels, channels, 1))
def forward(self, x):
return checkpoint(self._forward, (x,), self.parameters(), True) # TODO: check checkpoint usage, is True # TODO: fix the .half call!!!
def _forward(self, x):
b, c, *spatial = x.shape
x = x.reshape(b, c, -1)
qkv = self.qkv(self.norm(x))
h = self.attention(qkv)
h = self.proj_out(h)
return (x + h).reshape(b, c, *spatial)
class EncoderUNetModelWT(nn.Module):
"""
The half UNet model with attention and timestep embedding.
For usage, see UNet.
"""
def __init__(
self,
in_channels,
model_channels,
out_channels,
num_res_blocks,
attention_resolutions,
dropout=0,
channel_mult=(1, 2, 4, 8),
conv_resample=True,
dims=2,
use_checkpoint=False,
use_fp16=False,
num_heads=4,
num_head_channels=-1,
num_heads_upsample=-1,
use_scale_shift_norm=False,
resblock_updown=False,
use_new_attention_order=False,
):
super().__init__()
if num_heads_upsample == -1:
num_heads_upsample = num_heads
self.in_channels = in_channels
self.model_channels = model_channels
self.out_channels = out_channels
self.num_res_blocks = num_res_blocks
self.attention_resolutions = attention_resolutions
self.dropout = dropout
self.channel_mult = channel_mult
self.conv_resample = conv_resample
self.use_checkpoint = use_checkpoint
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
time_embed_dim = model_channels * 4
self.time_embed = nn.Sequential(
linear(model_channels, time_embed_dim),
nn.SiLU(),
linear(time_embed_dim, time_embed_dim),
)
self.input_blocks = nn.ModuleList(
[
TimestepEmbedSequential(
conv_nd(dims, in_channels, model_channels, 3, padding=1)
)
]
)
self._feature_size = model_channels
input_block_chans = []
ch = model_channels
ds = 1
for level, mult in enumerate(channel_mult):
for _ in range(num_res_blocks):
layers = [
ResBlock(
ch,
time_embed_dim,
dropout,
out_channels=mult * model_channels,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
)
]
ch = mult * model_channels
if ds in attention_resolutions:
layers.append(
AttentionBlock(
ch,
use_checkpoint=use_checkpoint,
num_heads=num_heads,
num_head_channels=num_head_channels,
use_new_attention_order=use_new_attention_order,
)
)
self.input_blocks.append(TimestepEmbedSequential(*layers))
self._feature_size += ch
if level != len(channel_mult) - 1:
out_ch = ch
self.input_blocks.append(
TimestepEmbedSequential(
ResBlock(
ch,
time_embed_dim,
dropout,
out_channels=out_ch,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
down=True,
)
if resblock_updown
else Downsample(
ch, conv_resample, dims=dims, out_channels=out_ch
)
)
)
ch = out_ch
input_block_chans.append(ch)
ds *= 2
self._feature_size += ch
self.middle_block = TimestepEmbedSequential(
ResBlock(
ch,
time_embed_dim,
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=num_head_channels,
use_new_attention_order=use_new_attention_order,
),
ResBlock(
ch,
time_embed_dim,
dropout,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
),
)
input_block_chans.append(ch)
self._feature_size += ch
self.input_block_chans = input_block_chans
self.fea_tran = nn.ModuleList([])
for i in range(len(input_block_chans)):
self.fea_tran.append(
ResBlock(
input_block_chans[i],
time_embed_dim,
dropout,
out_channels=out_channels,
dims=dims,
use_checkpoint=use_checkpoint,
use_scale_shift_norm=use_scale_shift_norm,
)
)
@torch.no_grad()
def forward(self, x, timesteps):
"""
Apply the model to an input batch.
:param x: an [N x C x ...] Tensor of inputs.
:param timesteps: a 1-D batch of timesteps.
:return: an [N x K] Tensor of outputs.
"""
emb = self.time_embed(timestep_embedding(timesteps, self.model_channels))
result_list = []
results = {}
h = x.type(self.dtype)
for module in self.input_blocks:
last_h = h
h = module(h, emb)
if h.size(-1) != last_h.size(-1):
result_list.append(last_h)
h = self.middle_block(h, emb)
result_list.append(h)
assert len(result_list) == len(self.fea_tran)
for i in range(len(result_list)):
results[str(result_list[i].size(-1))] = self.fea_tran[i](result_list[i], emb)
return results
def load_from_dict(self, state_dict):
"""
Load model weights from a dictionary.
:param state_dict: a dict of parameters.
"""
filtered_dict = {}
for k, v in state_dict.items():
if k.startswith("structcond_stage_model."):
filtered_dict[k[len("structcond_stage_model.") :]] = v
self.load_state_dict(filtered_dict)
def build_unetwt() -> EncoderUNetModelWT:
"""
Build a model from a state dict.
:param state_dict: a dict of parameters.
:return: a nn.Module.
"""
# The settings is from official setting yaml file.
# https://github.com/IceClear/StableSR/blob/main/configs/stableSRNew/v2-finetune_text_T_512.yaml
model = EncoderUNetModelWT(
in_channels=4,
model_channels=256,
out_channels=256,
num_res_blocks=2,
attention_resolutions=[ 4, 2, 1 ],
dropout=0.0,
channel_mult=[1, 1, 2, 2],
conv_resample=True,
dims=2,
use_checkpoint=False,
use_fp16=False,
num_heads=4,
num_head_channels=-1,
num_heads_upsample=-1,
use_scale_shift_norm=False,
resblock_updown=False,
use_new_attention_order=False,
)
global attn_func
attn_func = get_attn_func()
return model
if __name__ == "__main__":
'''
Test the lr encoder model.
'''
path = '../models/stablesr_sd21.ckpt'
state_dict = torch.load(path)
for key in state_dict.keys():
print(key)
model = build_unetwt()
model.load_from_dict(state_dict)
model = model.cuda()
test_latent = torch.randn(1, 4, 64, 64).half().cuda()
test_timesteps = torch.tensor([0]).half().cuda()
with torch.no_grad():
test_result = model(test_latent, test_timesteps)
print(test_result.keys())

20
tools/extract_srmodule.py Normal file
View File

@ -0,0 +1,20 @@
'''
This script extracts the spade and structcond module from the official stablesr_000117.ckpt
'''
import torch
stablesr_path = 'models/stablesr_000117.ckpt'
with open(stablesr_path, 'rb') as f:
stablesr_ckpt = torch.load(f, map_location='cpu')
srmodule = {}
for k, v in stablesr_ckpt['state_dict'].items():
if 'spade' in k or 'structcond' in k:
srmodule[k] = v
# print(k)
# save
torch.save(srmodule, 'models/stablesr_sd21.ckpt')

20
tools/extract_vaecfw.py Normal file
View File

@ -0,0 +1,20 @@
import torch
vae_path = 'models/vqgan_cfw_00011.ckpt'
with open(vae_path, 'rb') as f:
vae_ckpt = torch.load(f, map_location='cpu')
prune_keys = []
for k, v in vae_ckpt['state_dict'].items():
if 'decoder.fusion_layer' in k:
prune_keys.append(k)
print(k)
vae_cfw = {}
for k in prune_keys:
vae_cfw[k] = vae_ckpt['state_dict'][k]
del vae_ckpt['state_dict'][k]
torch.save(vae_ckpt, 'models/vqgan_cfw_00011_vae_only.ckpt')
torch.save(vae_cfw, 'models/vqgan_cfw_00011_cfw_only.ckpt')