263 lines
12 KiB
Python
263 lines
12 KiB
Python
import torch
|
||
|
||
from modules import devices, extra_networks
|
||
from modules.shared import state
|
||
|
||
from tile_utils.typing import *
|
||
from tile_utils.utils import *
|
||
from tile_methods.abstractdiffusion import TiledDiffusion
|
||
|
||
|
||
class MultiDiffusion(TiledDiffusion):
|
||
"""
|
||
Multi-Diffusion Implementation
|
||
https://arxiv.org/abs/2302.08113
|
||
"""
|
||
|
||
def __init__(self, p:StableDiffusionProcessing, *args, **kwargs):
|
||
super().__init__(p, *args, **kwargs)
|
||
assert p.sampler_name != 'UniPC', 'MultiDiffusion is not compatible with UniPC!'
|
||
|
||
# For ddim sampler we need to cache the pred_x0
|
||
self.x_pred_buffer = None
|
||
|
||
def hook(self):
|
||
if self.is_kdiff:
|
||
# For K-Diffusion sampler with uniform prompt, we hijack into the inner model for simplicity
|
||
# Otherwise, the masked-redraw will break due to the init_latent
|
||
self.sampler: CFGDenoiser
|
||
self.sampler_forward = self.sampler.inner_model.forward
|
||
self.sampler.inner_model.forward = self.kdiff_forward
|
||
else:
|
||
self.sampler: VanillaStableDiffusionSampler
|
||
self.sampler_forward = self.sampler.orig_p_sample_ddim
|
||
self.sampler.orig_p_sample_ddim = self.ddim_forward
|
||
|
||
@staticmethod
|
||
def unhook():
|
||
# no need to unhook MultiDiffusion as it only hook the sampler,
|
||
# which will be destroyed after the painting is done
|
||
pass
|
||
|
||
def reset_buffer(self, x_in:Tensor):
|
||
super().reset_buffer(x_in)
|
||
|
||
# ddim needs to cache pred0
|
||
if self.is_ddim:
|
||
if self.x_pred_buffer is None:
|
||
self.x_pred_buffer = torch.zeros_like(x_in, device=x_in.device)
|
||
else:
|
||
self.x_pred_buffer.zero_()
|
||
|
||
@custom_bbox
|
||
def init_custom_bbox(self, *args):
|
||
super().init_custom_bbox(*args)
|
||
|
||
for bbox in self.custom_bboxes:
|
||
if bbox.blend_mode == BlendMode.BACKGROUND:
|
||
self.weights[bbox.slicer] += 1.0
|
||
|
||
''' ↓↓↓ kernel hijacks ↓↓↓ '''
|
||
|
||
def repeat_cond_dict(self, cond_input:CondDict, bboxes:List[CustomBBox]) -> CondDict:
|
||
cond = cond_input['c_crossattn'][0]
|
||
# repeat the condition on its first dim
|
||
cond_shape = cond.shape
|
||
cond = cond.repeat((len(bboxes),) + (1,) * (len(cond_shape) - 1))
|
||
image_cond = cond_input['c_concat'][0]
|
||
if image_cond.shape[2] == self.h and image_cond.shape[3] == self.w:
|
||
image_cond_list = []
|
||
for bbox in bboxes:
|
||
image_cond_list.append(image_cond[bbox.slicer])
|
||
image_cond_tile = torch.cat(image_cond_list, dim=0)
|
||
else:
|
||
image_cond_shape = image_cond.shape
|
||
image_cond_tile = image_cond.repeat((len(bboxes),) + (1,) * (len(image_cond_shape) - 1))
|
||
return {"c_crossattn": [cond], "c_concat": [image_cond_tile]}
|
||
|
||
@torch.no_grad()
|
||
@keep_signature
|
||
def kdiff_forward(self, x_in:Tensor, sigma_in:Tensor, cond:CondDict):
|
||
'''
|
||
This function hijacks `k_diffusion.external.CompVisDenoiser.forward()`
|
||
So its signature should be the same as the original function, especially the "cond" should be with exactly the same name
|
||
'''
|
||
|
||
assert CompVisDenoiser.forward
|
||
# x_in: [B, C=4, H=64, W=64]
|
||
# sigma_in： [1]
|
||
# cond['c_crossattn'][0]: [1, 77, 768]
|
||
def org_func(x):
|
||
return self.sampler_forward(x, sigma_in, cond=cond)
|
||
|
||
def repeat_func(x_tile, bboxes):
|
||
# For kdiff sampler, the dim 0 of input x_in is:
|
||
# = batch_size * (num_AND + 1) if not an edit model
|
||
# = batch_size * (num_AND + 2) otherwise
|
||
sigma_in_tile = sigma_in.repeat(len(bboxes))
|
||
new_cond = self.repeat_cond_dict(cond, bboxes)
|
||
x_tile_out = self.sampler_forward(x_tile, sigma_in_tile, cond=new_cond)
|
||
return x_tile_out
|
||
|
||
def custom_func(x, bbox_id, bbox):
|
||
return self.kdiff_custom_forward(x, sigma_in, cond, bbox_id, bbox, self.sampler_forward)
|
||
|
||
return self.sample_one_step(x_in, org_func, repeat_func, custom_func)
|
||
|
||
@torch.no_grad()
|
||
@keep_signature
|
||
def ddim_forward(self, x_in:Tensor, cond_in:CondDict, ts:Tensor, unconditional_conditioning:Tensor, *args, **kwargs):
|
||
'''
|
||
This function will replace the original p_sample_ddim function in ldm/diffusionmodels/ddim.py
|
||
So its signature should be the same as the original function,
|
||
Particularly, the unconditional_conditioning should be with exactly the same name
|
||
'''
|
||
|
||
assert VanillaStableDiffusionSampler.p_sample_ddim_hook
|
||
|
||
def org_func(x):
|
||
return self.sampler_forward(x, cond_in, ts, unconditional_conditioning=unconditional_conditioning, *args, **kwargs)
|
||
|
||
def repeat_func(x_tile, bboxes):
|
||
if isinstance(cond_in, dict):
|
||
ts_tile = ts.repeat(len(bboxes))
|
||
cond_tile = self.repeat_cond_dict(cond_in, bboxes)
|
||
ucond_tile = self.repeat_cond_dict(unconditional_conditioning, bboxes)
|
||
else:
|
||
ts_tile = ts.repeat(len(bboxes))
|
||
cond_shape = cond_in.shape
|
||
cond_tile = cond_in.repeat((len(bboxes),) + (1,) * (len(cond_shape) - 1))
|
||
ucond_shape = unconditional_conditioning.shape
|
||
ucond_tile = unconditional_conditioning.repeat((len(bboxes),) + (1,) * (len(ucond_shape) - 1))
|
||
x_tile_out, x_pred = self.sampler_forward(
|
||
x_tile, cond_tile, ts_tile,
|
||
unconditional_conditioning=ucond_tile,
|
||
*args, **kwargs)
|
||
return x_tile_out, x_pred
|
||
|
||
def custom_func(x, bbox_id:int, bbox:CustomBBox):
|
||
# before the final forward, we can set the control tensor
|
||
def forward_func(x, *args, **kwargs):
|
||
self.set_controlnet_tensors(bbox_id, 2*x.shape[0])
|
||
return self.sampler_forward(x, *args, **kwargs)
|
||
return self.ddim_custom_forward(x, cond_in, bbox, ts, forward_func, *args, **kwargs)
|
||
|
||
return self.sample_one_step(x_in, org_func, repeat_func, custom_func)
|
||
|
||
def sample_one_step(self, x_in:Tensor, org_func: Callable, repeat_func:Callable, custom_func:Callable):
|
||
'''
|
||
this method splits the whole latent and process in tiles
|
||
- x_in: current whole U-Net latent
|
||
- org_func: original forward function, when use highres
|
||
- denoise_func: one step denoiser for grid tile
|
||
- denoise_custom_func: one step denoiser for custom tile
|
||
'''
|
||
|
||
N, C, H, W = x_in.shape
|
||
if H != self.h or W != self.w:
|
||
self.reset_controlnet_tensors()
|
||
return org_func(x_in)
|
||
|
||
# clear buffer canvas
|
||
self.reset_buffer(x_in)
|
||
|
||
# Background sampling (grid bbox)
|
||
if self.draw_background:
|
||
for batch_id, bboxes in enumerate(self.batched_bboxes):
|
||
if state.interrupted: return x_in
|
||
|
||
# batching
|
||
x_tile_list = []
|
||
for bbox in bboxes:
|
||
x_tile_list.append(x_in[bbox.slicer])
|
||
x_tile = torch.cat(x_tile_list, dim=0)
|
||
|
||
# controlnet tiling
|
||
# FIXME: is_denoise is default to False, however it is set to True in case of MixtureOfDiffusers
|
||
self.switch_controlnet_tensors(batch_id, N, len(bboxes))
|
||
|
||
# compute tiles
|
||
if self.is_kdiff:
|
||
x_tile_out = repeat_func(x_tile, bboxes)
|
||
for i, bbox in enumerate(bboxes):
|
||
self.x_buffer[bbox.slicer] += x_tile_out[i*N:(i+1)*N, :, :, :]
|
||
else:
|
||
x_tile_out, x_tile_pred = repeat_func(x_tile, bboxes)
|
||
for i, bbox in enumerate(bboxes):
|
||
self.x_buffer [bbox.slicer] += x_tile_out [i*N:(i+1)*N, :, :, :]
|
||
self.x_pred_buffer[bbox.slicer] += x_tile_pred[i*N:(i+1)*N, :, :, :]
|
||
|
||
# update progress bar
|
||
self.update_pbar()
|
||
|
||
# Custom region sampling (custom bbox)
|
||
x_feather_buffer = None
|
||
x_feather_mask = None
|
||
x_feather_count = None
|
||
x_feather_pred_buffer = None
|
||
if len(self.custom_bboxes) > 0:
|
||
for bbox_id, bbox in enumerate(self.custom_bboxes):
|
||
if state.interrupted: return x_in
|
||
|
||
if not self.p.disable_extra_networks:
|
||
with devices.autocast():
|
||
extra_networks.activate(self.p, bbox.extra_network_data)
|
||
|
||
x_tile = x_in[bbox.slicer]
|
||
|
||
if self.is_kdiff:
|
||
# retrieve original x_in from construncted input
|
||
x_tile_out = custom_func(x_tile, bbox_id, bbox)
|
||
|
||
if bbox.blend_mode == BlendMode.BACKGROUND:
|
||
self.x_buffer[bbox.slicer] += x_tile_out
|
||
elif bbox.blend_mode == BlendMode.FOREGROUND:
|
||
if x_feather_buffer is None:
|
||
x_feather_buffer = torch.zeros_like(self.x_buffer)
|
||
x_feather_mask = torch.zeros_like(self.x_buffer)
|
||
x_feather_count = torch.zeros_like(self.x_buffer)
|
||
x_feather_buffer[bbox.slicer] += x_tile_out
|
||
x_feather_mask [bbox.slicer] += bbox.feather_mask
|
||
x_feather_count [bbox.slicer] += 1
|
||
else:
|
||
x_tile_out, x_tile_pred = custom_func(x_tile, bbox_id, bbox)
|
||
|
||
if bbox.blend_mode == BlendMode.BACKGROUND:
|
||
self.x_buffer [bbox.slicer] += x_tile_out
|
||
self.x_pred_buffer[bbox.slicer] += x_tile_pred
|
||
elif bbox.blend_mode == BlendMode.FOREGROUND:
|
||
if x_feather_buffer is None:
|
||
x_feather_buffer = torch.zeros_like(self.x_buffer)
|
||
x_feather_mask = torch.zeros_like(self.x_buffer)
|
||
x_feather_count = torch.zeros_like(self.x_buffer)
|
||
x_feather_pred_buffer = torch.zeros_like(self.x_pred_buffer)
|
||
x_feather_buffer [bbox.slicer] += x_tile_out
|
||
x_feather_mask [bbox.slicer] += bbox.feather_mask
|
||
x_feather_count [bbox.slicer] += 1
|
||
x_feather_pred_buffer[bbox.slicer] += x_tile_pred
|
||
|
||
if not self.p.disable_extra_networks:
|
||
with devices.autocast():
|
||
extra_networks.deactivate(self.p, bbox.extra_network_data)
|
||
|
||
# update progress bar
|
||
self.update_pbar()
|
||
|
||
# Averaging background buffer
|
||
x_out = torch.where(self.weights > 1, self.x_buffer / self.weights, self.x_buffer)
|
||
if self.is_ddim:
|
||
x_pred_out = torch.where(self.weights > 1, self.x_pred_buffer / self.weights, self.x_pred_buffer)
|
||
|
||
# Foreground Feather blending
|
||
if x_feather_buffer is not None:
|
||
# Average overlapping feathered regions
|
||
x_feather_buffer = torch.where(x_feather_count > 1, x_feather_buffer / x_feather_count, x_feather_buffer)
|
||
x_feather_mask = torch.where(x_feather_count > 1, x_feather_mask / x_feather_count, x_feather_mask)
|
||
# Weighted average with original x_buffer
|
||
x_out = torch.where(x_feather_count > 0, x_out * (1 - x_feather_mask) + x_feather_buffer * x_feather_mask, x_out)
|
||
if self.is_ddim:
|
||
x_feather_pred_buffer = torch.where(x_feather_count > 1, x_feather_pred_buffer / x_feather_count, x_feather_pred_buffer)
|
||
x_pred_out = torch.where(x_feather_count > 0, x_pred_out * (1 - x_feather_mask) + x_feather_pred_buffer * x_feather_mask, x_pred_out)
|
||
|
||
return x_out if self.is_kdiff else (x_out, x_pred_out)
|