244 lines
11 KiB
Python
244 lines
11 KiB
Python
from tile_methods.abstractdiffusion import AbstractDiffusion
|
|
from tile_utils.utils import *
|
|
|
|
|
|
class MultiDiffusion(AbstractDiffusion):
|
|
"""
|
|
Multi-Diffusion Implementation
|
|
https://arxiv.org/abs/2302.08113
|
|
"""
|
|
|
|
def __init__(self, p:Processing, *args, **kwargs):
|
|
super().__init__(p, *args, **kwargs)
|
|
assert p.sampler_name != 'UniPC', 'MultiDiffusion is not compatible with UniPC!'
|
|
|
|
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: KDiffusionSampler
|
|
self.sampler.model_wrap_cfg: CFGDenoiserKDiffusion
|
|
self.sampler.model_wrap_cfg.inner_model: Union[CompVisDenoiser, CompVisVDenoiser]
|
|
self.sampler_forward = self.sampler.model_wrap_cfg.inner_model.forward
|
|
self.sampler.model_wrap_cfg.inner_model.forward = self.kdiff_forward
|
|
else:
|
|
self.sampler: CompVisSampler
|
|
self.sampler.model_wrap_cfg: CFGDenoiserTimesteps
|
|
self.sampler.model_wrap_cfg.inner_model: Union[CompVisTimestepsDenoiser, CompVisTimestepsVDenoiser]
|
|
self.sampler_forward = self.sampler.model_wrap_cfg.inner_model.forward
|
|
self.sampler.model_wrap_cfg.inner_model.forward = 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)
|
|
|
|
@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 ↓↓↓ '''
|
|
|
|
@torch.no_grad()
|
|
@keep_signature
|
|
def kdiff_forward(self, x_in:Tensor, sigma_in:Tensor, cond:CondDict) -> Tensor:
|
|
assert CompVisDenoiser.forward
|
|
assert CompVisVDenoiser.forward
|
|
|
|
def org_func(x:Tensor) -> Tensor:
|
|
return self.sampler_forward(x, sigma_in, cond=cond)
|
|
|
|
def repeat_func(x_tile:Tensor, bboxes:List[CustomBBox]) -> Tensor:
|
|
# 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_tile = self.repeat_tensor(sigma_in, len(bboxes))
|
|
cond_tile = self.repeat_cond_dict(cond, bboxes)
|
|
return self.sampler_forward(x_tile, sigma_tile, cond=cond_tile)
|
|
|
|
def custom_func(x:Tensor, bbox_id:int, bbox:CustomBBox) -> Tensor:
|
|
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, ts_in:Tensor, cond:Union[CondDict, Tensor]) -> Tensor:
|
|
assert CompVisTimestepsDenoiser.forward
|
|
assert CompVisTimestepsVDenoiser.forward
|
|
|
|
def org_func(x:Tensor) -> Tensor:
|
|
return self.sampler_forward(x, ts_in, cond=cond)
|
|
|
|
def repeat_func(x_tile:Tensor, bboxes:List[CustomBBox]) -> Tuple[Tensor, Tensor]:
|
|
n_rep = len(bboxes)
|
|
ts_tile = self.repeat_tensor(ts_in, n_rep)
|
|
if isinstance(cond, dict): # FIXME: when will enter this branch?
|
|
cond_tile = self.repeat_cond_dict(cond, bboxes)
|
|
else:
|
|
cond_tile = self.repeat_tensor(cond, n_rep)
|
|
return self.sampler_forward(x_tile, ts_tile, cond=cond_tile)
|
|
|
|
def custom_func(x:Tensor, bbox_id:int, bbox:CustomBBox) -> Tensor:
|
|
# before the final forward, we can set the control tensor
|
|
def forward_func(x, *args, **kwargs):
|
|
self.set_custom_controlnet_tensors(bbox_id, 2*x.shape[0])
|
|
self.set_custom_stablesr_tensors(bbox_id)
|
|
return self.sampler_forward(x, *args, **kwargs)
|
|
return self.ddim_custom_forward(x, cond, bbox, ts_in, forward_func)
|
|
|
|
return self.sample_one_step(x_in, org_func, repeat_func, custom_func)
|
|
|
|
def repeat_tensor(self, x:Tensor, n:int) -> Tensor:
|
|
''' repeat the tensor on it's first dim '''
|
|
if n == 1: return x
|
|
B = x.shape[0]
|
|
r_dims = len(x.shape) - 1
|
|
if B == 1: # batch_size = 1 (not `tile_batch_size`)
|
|
shape = [n] + [-1] * r_dims # [N, -1, ...]
|
|
return x.expand(shape) # `expand` is much lighter than `tile`
|
|
else:
|
|
shape = [n] + [1] * r_dims # [N, 1, ...]
|
|
return x.repeat(shape)
|
|
|
|
def repeat_cond_dict(self, cond_in:CondDict, bboxes:List[CustomBBox]) -> CondDict:
|
|
''' repeat all tensors in cond_dict on it's first dim (for a batch of tiles), returns a new object '''
|
|
# n_repeat
|
|
n_rep = len(bboxes)
|
|
# txt cond
|
|
tcond = self.get_tcond(cond_in) # [B=1, L, D] => [B*N, L, D]
|
|
tcond = self.repeat_tensor(tcond, n_rep)
|
|
# img cond
|
|
icond = self.get_icond(cond_in)
|
|
if icond.shape[2:] == (self.h, self.w): # img2img, [B=1, C, H, W]
|
|
icond = torch.cat([icond[bbox.slicer] for bbox in bboxes], dim=0)
|
|
else: # txt2img, [B=1, C=5, H=1, W=1]
|
|
icond = self.repeat_tensor(icond, n_rep)
|
|
# vec cond (SDXL)
|
|
vcond = self.get_vcond(cond_in) # [B=1, D]
|
|
if vcond is not None:
|
|
vcond = self.repeat_tensor(vcond, n_rep) # [B*N, D]
|
|
return self.make_cond_dict(cond_in, tcond, icond, vcond)
|
|
|
|
def sample_one_step(self, x_in:Tensor, org_func:Callable, repeat_func:Callable, custom_func:Callable) -> Tensor:
|
|
'''
|
|
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
|
|
- repeat_func: one step denoiser for grid tile
|
|
- custom_func: one step denoiser for custom tile
|
|
'''
|
|
|
|
N, C, H, W = x_in.shape
|
|
if (H, W) != (self.h, self.w):
|
|
# We don't tile highres, let's just use the original org_func
|
|
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 = torch.cat([x_in[bbox.slicer] for bbox in bboxes], dim=0) # [TB, C, TH, TW]
|
|
|
|
# controlnet tiling
|
|
# FIXME: is_denoise is default to False, however it is set to True in case of MixtureOfDiffusers, why?
|
|
self.switch_controlnet_tensors(batch_id, N, len(bboxes))
|
|
|
|
# stablesr tiling
|
|
self.switch_stablesr_tensors(batch_id)
|
|
|
|
# compute tiles
|
|
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, :, :, :]
|
|
|
|
# update progress bar
|
|
self.update_pbar()
|
|
|
|
# Custom region sampling (custom bbox)
|
|
x_feather_buffer = None
|
|
x_feather_mask = None
|
|
x_feather_count = 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]
|
|
|
|
# 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((1, 1, H, W), device=x_in.device)
|
|
x_feather_count = torch.zeros((1, 1, H, W), device=x_in.device)
|
|
x_feather_buffer[bbox.slicer] += x_tile_out
|
|
x_feather_mask [bbox.slicer] += bbox.feather_mask
|
|
x_feather_count [bbox.slicer] += 1
|
|
|
|
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)
|
|
|
|
# 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)
|
|
|
|
return x_out
|
|
|
|
def get_noise(self, x_in:Tensor, sigma_in:Tensor, cond_in:Dict[str, Tensor], step:int) -> Tensor:
|
|
# NOTE: The following code is analytically wrong but aesthetically beautiful
|
|
cond_in_original = cond_in.copy()
|
|
|
|
def org_func(x:Tensor):
|
|
return shared.sd_model.apply_model(x, sigma_in, cond=cond_in_original)
|
|
|
|
def repeat_func(x_tile:Tensor, bboxes:List[CustomBBox]):
|
|
sigma_in_tile = sigma_in.repeat(len(bboxes))
|
|
cond_out = self.repeat_cond_dict(cond_in_original, bboxes)
|
|
x_tile_out = shared.sd_model.apply_model(x_tile, sigma_in_tile, cond=cond_out)
|
|
return x_tile_out
|
|
|
|
def custom_func(x:Tensor, bbox_id:int, bbox:CustomBBox):
|
|
# The negative prompt in custom bbox should not be used for noise inversion
|
|
# otherwise the result will be astonishingly bad.
|
|
tcond = Condition.reconstruct_cond(bbox.cond, step).unsqueeze_(0)
|
|
icond = self.get_icond(cond_in_original)
|
|
if icond.shape[2:] == (self.h, self.w):
|
|
icond = icond[bbox.slicer]
|
|
cond_out = self.make_cond_dict(cond_in, tcond, icond)
|
|
return shared.sd_model.apply_model(x, sigma_in, cond=cond_out)
|
|
|
|
return self.sample_one_step(x_in, org_func, repeat_func, custom_func)
|