mirror of https://github.com/vladmandic/automatic
278 lines
13 KiB
Python
278 lines
13 KiB
Python
"""
|
|
based on article by TimothyAlexisVass
|
|
https://huggingface.co/blog/TimothyAlexisVass/explaining-the-sdxl-latent-space
|
|
"""
|
|
|
|
import os
|
|
import torch
|
|
from modules import devices
|
|
from modules.logger import log
|
|
from modules.vae import sd_vae_taesd
|
|
|
|
|
|
debug_enabled = os.environ.get('SD_HDR_DEBUG', None) is not None
|
|
debug = log.trace if debug_enabled else lambda *args, **kwargs: None
|
|
debug('Trace: HDR')
|
|
|
|
|
|
def sharpen_tensor(tensor, ratio=0):
|
|
if ratio == 0:
|
|
return tensor
|
|
kernel = torch.ones((3, 3), dtype=tensor.dtype, device=tensor.device)
|
|
kernel[1, 1] = 5.0
|
|
kernel /= kernel.sum()
|
|
kernel = kernel.expand(tensor.shape[-3], 1, kernel.shape[0], kernel.shape[1])
|
|
result_tmp = torch.nn.functional.conv2d(tensor, kernel, groups=tensor.shape[-3])
|
|
result = tensor.clone()
|
|
result[..., 1:-1, 1:-1] = result_tmp
|
|
output = (1.0 + ratio) * tensor + (0 - ratio) * result
|
|
return soft_clamp_tensor(output, threshold=0.95)
|
|
|
|
|
|
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))
|
|
return tensor
|
|
|
|
|
|
def center_tensor(tensor, channel_shift=0.0, full_shift=0.0, offset=0.0):
|
|
if channel_shift == 0 and full_shift == 0 and offset == 0:
|
|
return tensor
|
|
tensor -= tensor.mean(dim=(-1, -2), keepdim=True) * channel_shift
|
|
tensor -= tensor.mean() * full_shift - offset
|
|
return tensor
|
|
|
|
|
|
def maximize_tensor(tensor, boundary=1.0):
|
|
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 *= normalization_factor
|
|
return tensor
|
|
|
|
|
|
def get_color(colorstr):
|
|
if not colorstr:
|
|
colorstr = "#000000"
|
|
rgb = torch.tensor(tuple(int(colorstr.lstrip('#')[i:i + 2], 16) for i in (0, 2, 4))).to(dtype=torch.float32)
|
|
rgb = (rgb / 255).unsqueeze(-1).unsqueeze(-1).repeat(1, 64, 64).to(dtype=devices.dtype, device=devices.device)
|
|
color = sd_vae_taesd.encode(rgb).squeeze(0)[0:3, 5, 5]
|
|
return color
|
|
|
|
|
|
def color_adjust(tensor, colorstr, ratio):
|
|
color = get_color(colorstr)
|
|
for i in range(3):
|
|
tensor[i] = center_tensor(tensor[i], full_shift=1, offset=color[i]*(ratio/2))
|
|
return tensor
|
|
|
|
|
|
def correction(p, timestep, latent, step=0):
|
|
total = getattr(p, 'correction_total_steps', 0)
|
|
if total > 0:
|
|
progress = step / total # 0.0 = first step, ~1.0 = last step
|
|
is_early = progress < 0.05
|
|
is_mid = 0.2 <= progress <= 0.7
|
|
is_late = progress >= 0.8
|
|
is_sharpen = progress >= 0.7
|
|
is_very_late = progress >= 0.9
|
|
else:
|
|
# fallback to timestep-based ranges for non-flow-match schedulers
|
|
is_early = timestep > 950
|
|
is_mid = 600 < timestep < 900
|
|
is_late = timestep < 200
|
|
is_sharpen = timestep < 350
|
|
is_very_late = 1 < timestep < 100
|
|
if is_early and p.hdr_clamp:
|
|
latent = soft_clamp_tensor(latent, threshold=p.hdr_threshold, boundary=p.hdr_boundary)
|
|
p.extra_generation_params["Latent clamp"] = f'{p.hdr_threshold}/{p.hdr_boundary}'
|
|
if is_mid and p.hdr_color != 0:
|
|
n = getattr(p, 'correction_steps_mid', 1)
|
|
num_channels = latent.shape[0]
|
|
if num_channels <= 4:
|
|
# SDXL-style: channel 0 is brightness, channels 1+ are color
|
|
latent[1:] = center_tensor(latent[1:], channel_shift=p.hdr_color / n, full_shift=float(p.hdr_mode))
|
|
else:
|
|
# Multi-channel latents (Flux 2, etc.): apply to all channels
|
|
latent = center_tensor(latent, channel_shift=p.hdr_color / n, full_shift=float(p.hdr_mode))
|
|
p.extra_generation_params["Latent color"] = f'{p.hdr_color}'
|
|
if is_mid and p.hdr_tint_ratio != 0:
|
|
n = getattr(p, 'correction_steps_mid', 1)
|
|
num_channels = latent.shape[0]
|
|
if num_channels <= 4:
|
|
# SDXL-style: TAESD color encoding maps to 4-channel latent space
|
|
latent = color_adjust(latent, p.hdr_color_picker, p.hdr_tint_ratio / n)
|
|
else:
|
|
# Multi-channel latents: apply uniform offset to all channels based on tint ratio
|
|
latent = center_tensor(latent, full_shift=1.0, offset=p.hdr_tint_ratio / n)
|
|
p.extra_generation_params["Latent tint"] = f'{p.hdr_tint_ratio}'
|
|
p.extra_generation_params["Latent tint color"] = p.hdr_color_picker
|
|
if is_late and p.hdr_brightness != 0:
|
|
n = getattr(p, 'correction_steps_late', 1)
|
|
num_channels = latent.shape[0]
|
|
if num_channels <= 4:
|
|
# SDXL-style: brightness is in channel 0 (luminance)
|
|
latent[0:1] = center_tensor(latent[0:1], full_shift=float(p.hdr_mode), offset=p.hdr_brightness / n)
|
|
else:
|
|
# Multi-channel latents (Flux 2, etc.): scale intensity to avoid color shifts
|
|
scale = 1.0 + (p.hdr_brightness / n) * 0.25
|
|
latent = latent * scale
|
|
p.extra_generation_params["Latent brightness"] = f'{p.hdr_brightness}'
|
|
if is_sharpen and p.hdr_sharpen != 0:
|
|
progress_in_range = (step - int(total * 0.7)) / max(int(total * 0.3), 1) if total > 0 else timestep / 350
|
|
per_step_ratio = 2 ** (progress_in_range * 1.4) * p.hdr_sharpen / 16
|
|
if abs(per_step_ratio) > 0.01:
|
|
latent = sharpen_tensor(latent, ratio=per_step_ratio)
|
|
p.extra_generation_params["Latent sharpen"] = f'{p.hdr_sharpen}'
|
|
if is_very_late and p.hdr_maximize:
|
|
latent = center_tensor(latent, channel_shift=p.hdr_max_center, full_shift=1.0)
|
|
latent = maximize_tensor(latent, boundary=p.hdr_max_boundary)
|
|
p.extra_generation_params["Latent max"] = f'{p.hdr_max_center}/{p.hdr_max_boundary}'
|
|
return latent
|
|
|
|
|
|
def _unpack_latents(latents, pipe, p):
|
|
"""Unpack packed latents to standard [B, C, H, W] format for correction."""
|
|
vae_scale = getattr(pipe, 'vae_scale_factor', 8)
|
|
if p.hr_resize_mode > 0 and (p.hr_upscaler != 'None' or p.hr_resize_mode == 5) and p.is_hr_pass:
|
|
width = max(getattr(p, 'width', 0), getattr(p, 'hr_upscale_to_x', 0))
|
|
height = max(getattr(p, 'height', 0), getattr(p, 'hr_upscale_to_y', 0))
|
|
else:
|
|
width = getattr(p, 'width', 1024)
|
|
height = getattr(p, 'height', 1024)
|
|
if hasattr(pipe, '_unpack_latents') and hasattr(pipe, 'vae_scale_factor'):
|
|
# Flux 1 / Bria: use pipeline's own unpack method
|
|
unpacked = pipe._unpack_latents(latents, height, width, vae_scale) # pylint: disable=protected-access
|
|
return unpacked, 'flux1'
|
|
if hasattr(pipe, '_unpatchify_latents'):
|
|
# Flux 2: manual reshape [B, seq_len, patch_channels] -> [B, C, H, W]
|
|
b, seq_len, patch_ch = latents.shape
|
|
channels = patch_ch // 4
|
|
h_patches = height // vae_scale // 2
|
|
w_patches = width // vae_scale // 2
|
|
if h_patches * w_patches != seq_len:
|
|
h_patches = w_patches = int(seq_len ** 0.5)
|
|
unpacked = latents.view(b, h_patches, w_patches, channels, 2, 2)
|
|
unpacked = unpacked.permute(0, 3, 1, 4, 2, 5).reshape(b, channels, h_patches * 2, w_patches * 2)
|
|
return unpacked, 'flux2'
|
|
return latents, 'unknown'
|
|
|
|
|
|
def _repack_latents(latents, pack_type, pipe, p):
|
|
"""Repack standard [B, C, H, W] latents back to packed format."""
|
|
if p.hr_resize_mode > 0 and (p.hr_upscaler != 'None' or p.hr_resize_mode == 5) and p.is_hr_pass:
|
|
height = max(getattr(p, 'height', 0), getattr(p, 'hr_upscale_to_y', 0))
|
|
width = max(getattr(p, 'width', 0), getattr(p, 'hr_upscale_to_x', 0))
|
|
else:
|
|
height = getattr(p, 'height', 1024)
|
|
width = getattr(p, 'width', 1024)
|
|
if pack_type == 'flux1':
|
|
# Flux 1 / Bria: use pipeline's pack method
|
|
return pipe._pack_latents(latents, latents.shape[0], latents.shape[1], height, width) # pylint: disable=protected-access
|
|
if pack_type == 'flux2':
|
|
# Flux 2: manual repack [B, C, H, W] -> [B, seq_len, patch_channels]
|
|
b, channels, h, w = latents.shape
|
|
h_patches = h // 2
|
|
w_patches = w // 2
|
|
latents = latents.reshape(b, channels, h_patches, 2, w_patches, 2)
|
|
latents = latents.permute(0, 2, 4, 1, 3, 5).reshape(b, h_patches * w_patches, channels * 4)
|
|
return latents
|
|
return latents
|
|
|
|
|
|
def _count_steps_in_range(pipe, low, high):
|
|
"""Count scheduler timesteps that fall within (low, high) exclusive."""
|
|
timesteps = getattr(getattr(pipe, 'scheduler', None), 'timesteps', None)
|
|
if timesteps is None:
|
|
return 1
|
|
count = sum(1 for t in timesteps.tolist() if low < t < high)
|
|
return max(count, 1)
|
|
|
|
|
|
def _count_steps_below(pipe, threshold):
|
|
"""Count scheduler timesteps below a threshold."""
|
|
timesteps = getattr(getattr(pipe, 'scheduler', None), 'timesteps', None)
|
|
if timesteps is None:
|
|
return 1
|
|
count = sum(1 for t in timesteps.tolist() if t < threshold)
|
|
return max(count, 1)
|
|
|
|
|
|
def correction_callback(p, timestep, kwargs, pipe=None, initial: bool = False, step: int = 0):
|
|
if initial:
|
|
if not any([p.hdr_clamp, p.hdr_mode, p.hdr_maximize, p.hdr_sharpen, p.hdr_color, p.hdr_brightness, p.hdr_tint_ratio]):
|
|
p.correction_skip = True
|
|
return kwargs
|
|
# always skip for detailer passes (already-corrected image, different resolution)
|
|
if getattr(p, 'recursion', False):
|
|
p.correction_skip = True
|
|
return kwargs
|
|
# optionally skip for hires pass
|
|
if getattr(p, 'is_hr_pass', False) and not getattr(p, 'hdr_apply_hires', True):
|
|
p.correction_skip = True
|
|
return kwargs
|
|
p.correction_skip = False
|
|
p.correction_warned = False
|
|
total = getattr(pipe, 'num_timesteps', 0) if pipe is not None else 0
|
|
if total > 0:
|
|
p.correction_total_steps = total
|
|
p.correction_steps_mid = max(int(total * 0.5), 1) # 20%-70% range
|
|
p.correction_steps_late = max(int(total * 0.2), 1) # last 20%
|
|
elif pipe is not None:
|
|
p.correction_total_steps = 0
|
|
p.correction_steps_mid = _count_steps_in_range(pipe, 600, 900)
|
|
p.correction_steps_late = _count_steps_below(pipe, 200)
|
|
elif getattr(p, 'correction_skip', False):
|
|
return kwargs
|
|
latents = kwargs["latents"]
|
|
if debug_enabled:
|
|
debug(f'Correction callback: step={step} timestep={timestep} latents_shape={latents.shape} total={getattr(p, "correction_total_steps", "unset")} skip={getattr(p, "correction_skip", "unset")}')
|
|
if len(latents.shape) <= 3: # packed latent
|
|
if pipe is None:
|
|
if not getattr(p, 'correction_warned', False):
|
|
log.warning(f'Latent correction: shape={latents.shape} packed latent but no pipe reference')
|
|
p.correction_warned = True
|
|
return kwargs
|
|
unpacked, pack_type = _unpack_latents(latents, pipe, p)
|
|
if pack_type == 'unknown':
|
|
if not getattr(p, 'correction_warned', False):
|
|
log.warning(f'Latent correction: shape={latents.shape} unknown packed format')
|
|
p.correction_warned = True
|
|
return kwargs
|
|
for i in range(unpacked.shape[0]):
|
|
unpacked[i] = correction(p, timestep, unpacked[i], step=step)
|
|
kwargs["latents"] = _repack_latents(unpacked, pack_type, pipe, p)
|
|
elif len(latents.shape) == 4: # standard batched latent
|
|
for i in range(latents.shape[0]):
|
|
latents[i] = correction(p, timestep, latents[i], step=step)
|
|
if debug_enabled:
|
|
debug(f"Full Mean: {latents[i].mean().item()}")
|
|
debug(f"Channel Means: {latents[i].mean(dim=(-1, -2), keepdim=True).flatten().float().cpu().numpy()}")
|
|
debug(f"Channel Mins: {latents[i].min(-1, keepdim=True)[0].min(-2, keepdim=True)[0].flatten().float().cpu().numpy()}")
|
|
debug(f"Channel Maxes: {latents[i].max(-1, keepdim=True)[0].min(-2, keepdim=True)[0].flatten().float().cpu().numpy()}")
|
|
kwargs["latents"] = latents
|
|
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], step=step)
|
|
latents = latents.permute(1, 0, 2, 3).unsqueeze(0)
|
|
kwargs["latents"] = latents
|
|
else:
|
|
if not getattr(p, 'correction_warned', False):
|
|
log.warning(f'Latent correction: shape={latents.shape} unknown latent')
|
|
p.correction_warned = True
|
|
return kwargs
|