sd-webui-controlnet/scripts/hook.py

439 lines
18 KiB
Python

import torch
import einops
import torch.nn as nn
from enum import Enum
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
from ldm.modules.attention import BasicTransformerBlock
class ControlModelType(Enum):
"""
The type of Control Models (supported or not).
"""
ControlNet = "ControlNet, Lvmin Zhang"
T2I_Adapter = "T2I_Adapter, Chong Mou"
T2I_StyleAdapter = "T2I_StyleAdapter, Chong Mou"
T2I_CoAdapter = "T2I_CoAdapter, Chong Mou"
MasaCtrl = "MasaCtrl, Mingdeng Cao"
GLIGEN = "GLIGEN, Yuheng Li"
AttentionInjection = "AttentionInjection, Lvmin Zhang" # A simple attention injection written by Lvmin
StableSR = "StableSR, Jianyi Wang"
PromptDiffusion = "PromptDiffusion, Zhendong Wang"
ControlLoRA = "ControlLoRA, Wu Hecong"
# Written by Lvmin
class AttentionAutoMachine(Enum):
"""
Lvmin's algorithm for Attention AutoMachine States.
"""
Read = "Read"
Write = "Write"
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()
class ControlParams:
def __init__(
self,
control_model,
hint_cond,
weight,
guidance_stopped,
start_guidance_percent,
stop_guidance_percent,
advanced_weighting,
control_model_type,
hr_hint_cond,
global_average_pooling,
batch_size,
instance_counter,
is_vanilla_samplers,
cfg_scale,
soft_injection,
cfg_injection
):
self.control_model = control_model
self._hint_cond = hint_cond
self.weight = weight
self.guidance_stopped = guidance_stopped
self.start_guidance_percent = start_guidance_percent
self.stop_guidance_percent = stop_guidance_percent
self.advanced_weighting = advanced_weighting
self.control_model_type = control_model_type
self.global_average_pooling = global_average_pooling
self.hr_hint_cond = hr_hint_cond
self.used_hint_cond = None
self.used_hint_cond_latent = None
self.batch_size = batch_size
self.instance_counter = instance_counter
self.is_vanilla_samplers = is_vanilla_samplers
self.cfg_scale = cfg_scale
self.soft_injection = soft_injection
self.cfg_injection = cfg_injection
def generate_uc_mask(self, length, dtype=None, device=None, python_list=False):
if self.is_vanilla_samplers and self.cfg_scale == 1:
if python_list:
return [1 for _ in range(length)]
return torch.tensor([1 for _ in range(length)], dtype=dtype, device=device)
y = []
for i in range(length):
p = (self.instance_counter + i) % (self.batch_size * 2)
if self.is_vanilla_samplers:
y += [0] if p < self.batch_size else [1]
else:
y += [1] if p < self.batch_size else [0]
self.instance_counter += length
if python_list:
return y
return torch.tensor(y, dtype=dtype, device=device)
@property
def hint_cond(self):
return self._hint_cond
# fix for all the extensions that modify hint_cond,
# by forcing used_hint_cond to update on the next timestep
# hr_hint_cond can stay the same, since most extensions dont modify the hires pass
# but if they do, it will cause problems
@hint_cond.setter
def hint_cond(self, new_hint_cond):
self._hint_cond = new_hint_cond
self.used_hint_cond = None
self.used_hint_cond_latent = None
def aligned_adding(base, x, require_channel_alignment):
if isinstance(x, float):
if x == 0.0:
return base
return base + x
if require_channel_alignment:
zeros = torch.zeros_like(base)
zeros[:, :x.shape[1], ...] = x
x = zeros
# resize to sample resolution
base_h, base_w = base.shape[-2:]
xh, xw = x.shape[-2:]
if base_h != xh or base_w != xw:
x = th.nn.functional.interpolate(x, size=(base_h, base_w), mode="nearest")
return base + x
# DFS Search for Torch.nn.Module, Written by Lvmin
def torch_dfs(model: torch.nn.Module):
result = [model]
for child in model.children():
result += torch_dfs(child)
return result
class UnetHook(nn.Module):
def __init__(self, lowvram=False) -> None:
super().__init__()
self.lowvram = lowvram
self.model = None
self.sd_ldm = None
self.control_params = None
self.attention_auto_machine = AttentionAutoMachine.Read
self.attention_auto_machine_uc_mask = None
self.attention_auto_machine_weight = 1.0
def guidance_schedule_handler(self, x):
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
def hook(self, model, sd_ldm, control_params):
self.model = model
self.sd_ldm = sd_ldm
self.control_params = control_params
outer = self
def forward(self, x, timesteps=None, context=None, **kwargs):
total_controlnet_embedding = [0.0] * 13
total_t2i_adapter_embedding = [0.0] * 4
require_inpaint_hijack = False
is_in_high_res_fix = False
# High-res fix
for param in outer.control_params:
# select which hint_cond to use
param.used_hint_cond = param.hint_cond
# Attention Injection do not need high-res fix
if param.control_model_type in [ControlModelType.AttentionInjection]:
continue
# has high-res fix
if param.hr_hint_cond is not None and x.ndim == 4 and param.hint_cond.ndim == 3 and param.hr_hint_cond.ndim == 3:
_, h_lr, w_lr = param.hint_cond.shape
_, h_hr, w_hr = param.hr_hint_cond.shape
_, _, h, w = x.shape
h, w = h * 8, w * 8
if abs(h - h_lr) < abs(h - h_hr):
# we are in low-res path
param.used_hint_cond = param.hint_cond
else:
# we are in high-res path
param.used_hint_cond = param.hr_hint_cond
is_in_high_res_fix = True
# Convert control image to latent
for param in outer.control_params:
if param.used_hint_cond_latent is not None:
continue
if param.control_model_type not in [ControlModelType.AttentionInjection]:
continue
query_size = int(x.shape[0])
latent_hint = param.used_hint_cond[None] * 2.0 - 1.0
latent_hint = outer.sd_ldm.encode_first_stage(latent_hint)
latent_hint = outer.sd_ldm.get_first_stage_encoding(latent_hint)
latent_hint = torch.cat([latent_hint.clone() for _ in range(query_size)], dim=0)
param.used_hint_cond_latent = latent_hint
# handle prompt token control
for param in outer.control_params:
if param.guidance_stopped:
continue
if param.control_model_type not in [ControlModelType.T2I_StyleAdapter]:
continue
param.control_model.to(devices.get_device_for("controlnet"))
query_size = int(x.shape[0])
control = param.control_model(x=x, hint=param.used_hint_cond, timesteps=timesteps, context=context)
uc_mask = param.generate_uc_mask(query_size, dtype=x.dtype, device=x.device)[:, None, None]
control = torch.cat([control.clone() for _ in range(query_size)], dim=0)
control *= param.weight
control *= uc_mask
context = torch.cat([context, control.clone()], dim=1)
# handle ControlNet / T2I_Adapter
for param in outer.control_params:
if param.guidance_stopped:
continue
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
if param.control_model_type == ControlModelType.ControlNet:
if x.shape[1] != control_model.input_blocks[0][0].in_channels and x.shape[1] == 9:
# inpaint_model: 4 data + 4 downscaled image + 1 mask
x_in = x[:, :4, ...]
require_inpaint_hijack = True
assert param.used_hint_cond is not None, f"Controlnet is enabled but no input image is given"
control = param.control_model(x=x_in, hint=param.used_hint_cond, timesteps=timesteps, context=context)
control_scales = ([param.weight] * 13)
if outer.lowvram:
param.control_model.to("cpu")
if param.cfg_injection or param.global_average_pooling:
query_size = int(x.shape[0])
if param.control_model_type == ControlModelType.T2I_Adapter:
control = [torch.cat([c.clone() for _ in range(query_size)], dim=0) for c in control]
uc_mask = param.generate_uc_mask(query_size, dtype=x.dtype, device=x.device)[:, None, None, None]
control = [c * uc_mask for c in control]
if param.soft_injection or is_in_high_res_fix:
# important! use the soft weights with high-res fix can significantly reduce artifacts.
if param.control_model_type == ControlModelType.T2I_Adapter:
control_scales = [param.weight * x for x in (0.25, 0.62, 0.825, 1.0)]
elif param.control_model_type == ControlModelType.ControlNet:
control_scales = [param.weight * (0.825 ** float(12 - i)) for i in range(13)]
if param.advanced_weighting is not None:
control_scales = param.advanced_weighting
control = [c * scale for c, scale in zip(control, control_scales)]
if param.global_average_pooling:
control = [torch.mean(c, dim=(2, 3), keepdim=True) for c in control]
for idx, item in enumerate(control):
target = None
if param.control_model_type == ControlModelType.ControlNet:
target = total_controlnet_embedding
if param.control_model_type == ControlModelType.T2I_Adapter:
target = total_t2i_adapter_embedding
if target is not None:
target[idx] = item + target[idx]
# Handle attention-based control
for param in outer.control_params:
if param.guidance_stopped:
continue
if param.used_hint_cond_latent is None:
continue
if param.control_model_type not in [ControlModelType.AttentionInjection]:
continue
query_size = int(x.shape[0])
ref_xt = outer.sd_ldm.q_sample(param.used_hint_cond_latent, torch.round(timesteps).long())
outer.attention_auto_machine_uc_mask = param.generate_uc_mask(query_size, python_list=True)
if param.soft_injection:
outer.attention_auto_machine_uc_mask = [1 for _ in outer.attention_auto_machine_uc_mask]
outer.attention_auto_machine_weight = param.weight
outer.attention_auto_machine = AttentionAutoMachine.Write
outer.original_forward(x=ref_xt, timesteps=timesteps, context=context)
outer.attention_auto_machine = AttentionAutoMachine.Read
# U-Net Encoder
hs = []
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)
for i, module in enumerate(self.input_blocks):
h = module(h, emb, context)
if (i + 1) % 3 == 0:
h = aligned_adding(h, total_t2i_adapter_embedding.pop(0), require_inpaint_hijack)
hs.append(h)
h = self.middle_block(h, emb, context)
# U-Net Middle Block
h = aligned_adding(h, total_controlnet_embedding.pop(), 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)
h = module(h, emb, context)
# U-Net Output
h = h.type(x.dtype)
h = self.out(h)
return h
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)
finally:
if self.lowvram:
for param in self.control_params:
if param.control_model is not None:
param.control_model.to("cpu")
def hacked_basic_transformer_inner_forward(self, x, context=None):
x_norm1 = self.norm1(x)
self_attn1 = 0
if self.disable_self_attn:
# Do not use self-attention
self_attn1 = self.attn1(x_norm1, context=context)
else:
# Use self-attention
self_attention_context = x_norm1
if outer.attention_auto_machine == AttentionAutoMachine.Write:
uc_mask = outer.attention_auto_machine_uc_mask
control_weight = outer.attention_auto_machine_weight
store = []
for i, mask in enumerate(uc_mask):
if mask > 0.5 and control_weight > self.attn_weight:
store.append(self_attention_context[i])
else:
store.append(None)
self.bank.append(store)
self_attn1 = self.attn1(x_norm1, context=self_attention_context)
if outer.attention_auto_machine == AttentionAutoMachine.Read:
query_size = self_attention_context.shape[0]
self_attention_context = [self_attention_context[i] for i in range(query_size)]
for store in self.bank:
for i, v in enumerate(store):
if v is not None:
self_attention_context[i] = torch.cat([self_attention_context[i], v], dim=0)
x_norm1 = [x_norm1[i] for i in range(query_size)]
self_attn1 = [self.attn1(a[None], context=b[None]) for a, b in zip(x_norm1, self_attention_context)]
self_attn1 = torch.cat(self_attn1, dim=0)
self.bank.clear()
x = self_attn1 + x
x = self.attn2(self.norm2(x), context=context) + x
x = self.ff(self.norm3(x)) + x
return x
model._original_forward = model.forward
outer.original_forward = model.forward
model.forward = forward_webui.__get__(model, UNetModel)
attn_modules = [module for module in torch_dfs(model) if isinstance(module, BasicTransformerBlock)]
attn_modules = sorted(attn_modules, key=lambda x: - x.norm1.normalized_shape[0])
for i, module in enumerate(attn_modules):
module._original_inner_forward = module._forward
module._forward = hacked_basic_transformer_inner_forward.__get__(module, BasicTransformerBlock)
module.bank = []
module.attn_weight = float(i) / float(len(attn_modules))
scripts.script_callbacks.on_cfg_denoiser(self.guidance_schedule_handler)
def restore(self, model):
scripts.script_callbacks.remove_callbacks_for_function(self.guidance_schedule_handler)
if hasattr(self, "control_params"):
del self.control_params
if not hasattr(model, "_original_forward"):
# no such handle, ignore
return
model.forward = model._original_forward
del model._original_forward