mirror of
https://github.com/vladmandic/sdnext.git
synced 2026-01-27 15:02:48 +03:00
89 lines
3.9 KiB
Python
89 lines
3.9 KiB
Python
"""
|
|
based on article by TimothyAlexisVass
|
|
https://huggingface.co/blog/TimothyAlexisVass/explaining-the-sdxl-latent-space
|
|
"""
|
|
|
|
import os
|
|
import torch
|
|
from modules import shared
|
|
|
|
|
|
debug = shared.log.trace if os.environ.get('SD_HDR_DEBUG', None) is not None else lambda *args, **kwargs: None
|
|
debug('Trace: HDR')
|
|
|
|
|
|
def soft_clamp_tensor(tensor, threshold=0.8, boundary=4):
|
|
# shrinking towards the mean; will also remove outliers
|
|
if max(abs(tensor.max()), abs(tensor.min())) < boundary or threshold == 0:
|
|
return tensor
|
|
channel_dim = 0
|
|
threshold *= boundary
|
|
max_vals = tensor.max(channel_dim, keepdim=True)[0]
|
|
max_replace = ((tensor - threshold) / (max_vals - threshold)) * (boundary - threshold) + threshold
|
|
over_mask = tensor > threshold
|
|
min_vals = tensor.min(channel_dim, keepdim=True)[0]
|
|
min_replace = ((tensor + threshold) / (min_vals + threshold)) * (-boundary + threshold) - threshold
|
|
under_mask = tensor < -threshold
|
|
tensor = torch.where(over_mask, max_replace, torch.where(under_mask, min_replace, tensor))
|
|
debug(f'HDR soft clamp: threshold={threshold} boundary={boundary} shape={tensor.shape}')
|
|
return tensor
|
|
|
|
|
|
def center_tensor(tensor, channel_shift=1.0, full_shift=1.0, channels=[0, 1, 2, 3]): # pylint: disable=dangerous-default-value # noqa: B006
|
|
if channel_shift == 0 and full_shift == 0:
|
|
return tensor
|
|
means = []
|
|
for channel in channels:
|
|
means.append(tensor[0, channel].mean())
|
|
# tensor[0, channel] -= means[-1] * channel_shift
|
|
tensor[channel] -= means[-1] * channel_shift
|
|
tensor = tensor - tensor.mean() * full_shift
|
|
debug(f'HDR center: channel-shift={channel_shift} full-shift={full_shift} means={torch.stack(means)} shape={tensor.shape}')
|
|
return tensor
|
|
|
|
|
|
def maximize_tensor(tensor, boundary=1.0, _channels=[0, 1, 2]): # pylint: disable=dangerous-default-value # noqa: B006
|
|
if boundary == 1.0:
|
|
return tensor
|
|
boundary *= 4
|
|
min_val = tensor.min()
|
|
max_val = tensor.max()
|
|
normalization_factor = boundary / max(abs(min_val), abs(max_val))
|
|
# tensor[0, channels] *= normalization_factor
|
|
tensor *= normalization_factor
|
|
debug(f'HDR maximize: boundary={boundary} min={min_val} max={max_val} factor={normalization_factor} shape={tensor.shape}')
|
|
return tensor
|
|
|
|
|
|
def correction(p, timestep, latent):
|
|
if timestep > 950 and p.hdr_clamp:
|
|
p.extra_generation_params["HDR clamp"] = f'{p.hdr_threshold}/{p.hdr_boundary}'
|
|
latent = soft_clamp_tensor(latent, threshold=p.hdr_threshold, boundary=p.hdr_boundary)
|
|
if timestep > 700 and p.hdr_center:
|
|
p.extra_generation_params["HDR center"] = f'{p.hdr_channel_shift}/{p.hdr_full_shift}'
|
|
latent = center_tensor(latent, channel_shift=p.hdr_channel_shift, full_shift=p.hdr_full_shift)
|
|
if timestep > 1 and timestep < 100 and p.hdr_maximize:
|
|
p.extra_generation_params["HDR max"] = f'{p.hdr_max_center}/{p.hdr_max_boundry}'
|
|
latent = center_tensor(latent, channel_shift=p.hdr_max_center, full_shift=1.0)
|
|
latent = maximize_tensor(latent, boundary=p.hdr_max_boundry)
|
|
return latent
|
|
|
|
|
|
def correction_callback(p, timestep, kwargs):
|
|
if not p.hdr_clamp and not p.hdr_center and not p.hdr_maximize:
|
|
return kwargs
|
|
latents = kwargs["latents"]
|
|
# debug(f'HDR correction: latents={latents.shape}')
|
|
if len(latents.shape) == 4: # standard batched latent
|
|
for i in range(latents.shape[0]):
|
|
latents[i] = correction(p, timestep, latents[i])
|
|
elif len(latents.shape) == 5 and latents.shape[0] == 1: # probably animatediff
|
|
latents = latents.squeeze(0).permute(1, 0, 2, 3)
|
|
for i in range(latents.shape[0]):
|
|
latents[i] = correction(p, timestep, latents[i])
|
|
latents = latents.permute(1, 0, 2, 3).unsqueeze(0)
|
|
else:
|
|
shared.log.debug(f'HDR correction: unknown latent shape {latents.shape}')
|
|
kwargs["latents"] = latents
|
|
return kwargs
|