Spaces:
Running on Zero
Running on Zero
File size: 14,872 Bytes
b701455 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 | """
Tiny AutoEncoder for Stable Diffusion
(DNN for encoding / decoding SD's latent space)
"""
# TODO: Check if multiprocessing is possible for this module
from PIL import Image
import numpy as np
import torch
from src.Utilities import util
import torch.nn as nn
from src.cond import cast
from src.user import app_instance
def conv(n_in: int, n_out: int, **kwargs) -> cast.disable_weight_init.Conv2d:
"""#### Create a convolutional layer.
#### Args:
- `n_in` (int): The number of input channels.
- `n_out` (int): The number of output channels.
#### Returns:
- `torch.nn.Module`: The convolutional layer.
"""
return cast.disable_weight_init.Conv2d(n_in, n_out, 3, padding=1, **kwargs)
class Clamp(nn.Module):
"""#### Class representing a clamping layer."""
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""#### Forward pass of the clamping layer.
#### Args:
- `x` (torch.Tensor): The input tensor.
#### Returns:
- `torch.Tensor`: The clamped tensor.
"""
return torch.tanh(x / 3) * 3
class Block(nn.Module):
"""#### Class representing a block layer."""
def __init__(self, n_in: int, n_out: int):
"""#### Initialize the block layer.
#### Args:
- `n_in` (int): The number of input channels.
- `n_out` (int): The number of output channels.
#### Returns:
- `Block`: The block layer.
"""
super().__init__()
self.conv = nn.Sequential(
conv(n_in, n_out),
nn.ReLU(),
conv(n_out, n_out),
nn.ReLU(),
conv(n_out, n_out),
)
self.skip = (
cast.disable_weight_init.Conv2d(n_in, n_out, 1, bias=False)
if n_in != n_out
else nn.Identity()
)
self.fuse = nn.ReLU()
def forward(self, x: torch.Tensor) -> torch.Tensor:
return self.fuse(self.conv(x) + self.skip(x))
def Encoder2(latent_channels: int = 4) -> nn.Sequential:
"""#### Create an encoder.
#### Args:
- `latent_channels` (int, optional): The number of latent channels. Defaults to 4.
#### Returns:
- `torch.nn.Module`: The encoder.
"""
return nn.Sequential(
conv(3, 64),
Block(64, 64),
conv(64, 64, stride=2, bias=False),
Block(64, 64),
Block(64, 64),
Block(64, 64),
conv(64, 64, stride=2, bias=False),
Block(64, 64),
Block(64, 64),
Block(64, 64),
conv(64, 64, stride=2, bias=False),
Block(64, 64),
Block(64, 64),
Block(64, 64),
conv(64, latent_channels),
)
def Decoder2(latent_channels: int = 4) -> nn.Sequential:
"""#### Create a decoder.
#### Args:
- `latent_channels` (int, optional): The number of latent channels. Defaults to 4.
#### Returns:
- `torch.nn.Module`: The decoder.
"""
return nn.Sequential(
Clamp(),
conv(latent_channels, 64),
nn.ReLU(),
Block(64, 64),
Block(64, 64),
Block(64, 64),
nn.Upsample(scale_factor=2),
conv(64, 64, bias=False),
Block(64, 64),
Block(64, 64),
Block(64, 64),
nn.Upsample(scale_factor=2),
conv(64, 64, bias=False),
Block(64, 64),
Block(64, 64),
Block(64, 64),
nn.Upsample(scale_factor=2),
conv(64, 64, bias=False),
Block(64, 64),
conv(64, 3),
)
class TAESD(nn.Module):
"""#### Class representing a Tiny AutoEncoder for Stable Diffusion.
#### Attributes:
- `latent_magnitude` (float): Magnitude of the latent space.
- `latent_shift` (float): Shift value for the latent space.
- `vae_shift` (torch.nn.Parameter): Shift parameter for the VAE.
- `vae_scale` (torch.nn.Parameter): Scale parameter for the VAE.
- `taesd_encoder` (Encoder2): Encoder network for the TAESD.
- `taesd_decoder` (Decoder2): Decoder network for the TAESD.
#### Args:
- `encoder_path` (str, optional): Path to the encoder model file. Defaults to None.
- `decoder_path` (str, optional): Path to the decoder model file. Defaults to "./include/vae_approx/taesd_decoder.safetensors".
- `latent_channels` (int, optional): Number of channels in the latent space. Defaults to 4.
#### Methods:
- `scale_latents(x)`:
Scales raw latents to the range [0, 1].
- `unscale_latents(x)`:
Unscales latents from the range [0, 1] to raw latents.
- `decode(x)`:
Decodes the given latent representation to the original space.
- `encode(x)`:
Encodes the given input to the latent space.
"""
latent_magnitude = 3
latent_shift = 0.5
def __init__(
self,
encoder_path: str = None,
decoder_path: str = None,
latent_channels: int = 4,
):
"""#### Initialize the TAESD model.
#### Args:
- `encoder_path` (str, optional): Path to the encoder model file. Defaults to None.
- `decoder_path` (str, optional): Path to the decoder model file. Defaults to "./include/vae_approx/taesd_decoder.safetensors".
- `latent_channels` (int, optional): Number of channels in the latent space. Defaults to 4.
"""
super().__init__()
# Use torch factories to keep trace graph free from torch.tensor constructors.
self.vae_shift = torch.nn.Parameter(torch.zeros(1))
self.vae_scale = torch.nn.Parameter(torch.ones(1))
self.taesd_encoder = Encoder2(latent_channels)
self.taesd_decoder = Decoder2(latent_channels)
if decoder_path is None:
if latent_channels == 16:
decoder_path = "./include/vae_approx/diffusion_pytorch_model.safetensors"
else:
decoder_path = "./include/vae_approx/taesd_decoder.safetensors"
if encoder_path is not None:
self.taesd_encoder.load_state_dict(
util.load_torch_file(encoder_path, safe_load=True)
)
if decoder_path is not None:
sd = util.load_torch_file(decoder_path, safe_load=True)
# Load top-level shift/scale parameters if they exist in the checkpoint
if "vae_shift" in sd:
self.vae_shift.data.copy_(sd["vae_shift"])
if "vae_scale" in sd:
self.vae_scale.data.copy_(sd["vae_scale"])
# Fix for Flux taef1 checkpoint structure
if any(k.startswith("decoder.layers.") for k in sd.keys()):
new_sd = {}
for k, v in sd.items():
if k.startswith("decoder.layers."):
# k is like "decoder.layers.0.weight"
parts = k.split(".")
# parts = ["decoder", "layers", "0", "weight"]
try:
idx = int(parts[2])
new_idx = idx + 1
new_key = f"{new_idx}." + ".".join(parts[3:])
new_sd[new_key] = v
except ValueError:
pass
self.taesd_decoder.load_state_dict(new_sd)
else:
# Filter out top-level parameters before loading into decoder
decoder_sd = {k: v for k, v in sd.items() if k not in ["vae_shift", "vae_scale"]}
self.taesd_decoder.load_state_dict(decoder_sd)
@staticmethod
def scale_latents(x: torch.Tensor) -> torch.Tensor:
"""#### Scales raw latents to the range [0, 1].
#### Args:
- `x` (torch.Tensor): The raw latents.
#### Returns:
- `torch.Tensor`: The scaled latents.
"""
return x.div(2 * TAESD.latent_magnitude).add(TAESD.latent_shift).clamp(0, 1)
@staticmethod
def unscale_latents(x: torch.Tensor) -> torch.Tensor:
"""#### Unscales latents from the range [0, 1] to raw latents.
#### Args:
- `x` (torch.Tensor): The scaled latents.
#### Returns:
- `torch.Tensor`: The raw latents.
"""
return x.sub(TAESD.latent_shift).mul(2 * TAESD.latent_magnitude)
def decode(self, x: torch.Tensor) -> torch.Tensor:
"""#### Decodes the given latent representation to the original space.
#### Args:
- `x` (torch.Tensor): The latent representation.
#### Returns:
- `torch.Tensor`: The decoded representation.
"""
device = next(self.taesd_decoder.parameters()).device
x = x.to(device)
x_sample = self.taesd_decoder((x - self.vae_shift) * self.vae_scale)
x_sample = x_sample.sub(0.5).mul(2)
return x_sample
def encode(self, x: torch.Tensor) -> torch.Tensor:
"""#### Encodes the given input to the latent space.
#### Args:
- `x` (torch.Tensor): The input.
#### Returns:
- `torch.Tensor`: The latent representation.
"""
device = next(self.taesd_encoder.parameters()).device
x = x.to(device)
x_sample = (x + 1) / 2
latent = self.taesd_encoder(x_sample)
latent = latent / self.vae_scale + self.vae_shift
return latent
return (self.taesd_encoder(x * 0.5 + 0.5) / self.vae_scale) + self.vae_shift
from src.Device.ModelCache import get_model_cache
def decode_latents_to_images(x: torch.Tensor, flux: bool = False) -> list[Image.Image]:
"""Decode latents to PIL images using TAESD or approximation.
Includes robustness checks for NaN/Inf and out-of-range values that
can otherwise cause black preview images.
"""
if x is None:
return []
# Robustness: Handle NaNs and extreme values that cause black images
if torch.isnan(x).any() or torch.isinf(x).any():
x = torch.nan_to_num(x, nan=0.0, posinf=10.0, neginf=-10.0)
latent_channels = x.shape[1]
cache = get_model_cache()
images = []
decoded_batch = None
# If we have TAESD model for these channels, use it (4 for SD, 16 for Flux1)
if latent_channels in (4, 16):
taesd_instance = cache.get_taesd(latent_channels, flux)
if taesd_instance is None:
taesd_instance = TAESD(latent_channels=latent_channels)
# Use same dtype as latents for efficiency
taesd_instance.to(x.device, dtype=x.dtype)
cache.cache_taesd(latent_channels, flux, taesd_instance)
elif next(taesd_instance.parameters()).device != x.device or next(taesd_instance.parameters()).dtype != x.dtype:
taesd_instance.to(x.device, dtype=x.dtype)
with torch.no_grad():
# Optimization for large batches: only preview up to 4 images
if x.shape[0] > 4:
x = x[:4]
# Robustness: Clamp latents to a reasonable range for the decoder
# Standard TAESD expects latents roughly in [-5, 5] to [-10, 10] range
x_clamped = torch.clamp(x, -12.0, 12.0)
decoded_batch = taesd_instance.decode(x_clamped)
# Normalize to [0, 1] range for both SD and Flux
# Note: No channel swap needed - TAESD outputs RGB correctly for all models
decoded_batch = decoded_batch.add(1.0).mul(0.5)
# Apply sRGB transfer to approximate final display appearance before
# converting to uint8. This helps previews better match final images.
try:
if getattr(app_instance.app, "preview_srgb", True):
from src.Utilities import color as color_utils
decoded_batch = color_utils.linear_to_srgb(decoded_batch)
except Exception:
# Non-fatal: fall back to simple clamp if anything goes wrong
decoded_batch = decoded_batch.clamp(0, 1)
finally:
decoded_batch = decoded_batch.clamp(0, 1)
# For Flux2 (32 channels), use RGB approximation since no TAESD exists for 32ch
elif latent_channels == 32:
try:
from src.Utilities import Latent
lf = Latent.Flux2()
factors = torch.tensor(lf.latent_rgb_factors, device=x.device, dtype=x.dtype)
bias = torch.tensor(lf.latent_rgb_factors_bias, device=x.device, dtype=x.dtype)
# Simple linear preview: [B, 32, H, W] @ [32, 3] -> [B, 3, H, W]
with torch.no_grad():
if x.shape[0] > 4:
x = x[:4]
# Permute for matmul: [B, H, W, 32] @ [32, 3]
x_perm = x.permute(0, 2, 3, 1)
decoded_batch = torch.matmul(x_perm, factors) + bias
# Back to [B, 3, H, W]
decoded_batch = decoded_batch.permute(0, 3, 1, 2).clamp(0, 1)
except Exception:
return []
else:
return [] # Unsupported channels
if decoded_batch is not None:
# Optimization: Use non_blocking=True for CPU transfer to avoid GPU stall
# Final safety: ensure no NaNs survived to this point (uint8 cast of NaN is 0)
decoded_batch = torch.nan_to_num(decoded_batch, nan=0.0)
decoded_np = (decoded_batch.mul(255.0).to("cpu", dtype=torch.uint8, non_blocking=True).numpy())
# Use simple transpose and PIL conversion
# decoded_np is [B, C, H, W] -> [B, H, W, C]
decoded_np = np.transpose(decoded_np, (0, 2, 3, 1))
for i in range(decoded_np.shape[0]):
img = Image.fromarray(decoded_np[i], mode='RGB')
# Reduce preview size for faster base64 conversion and lower bandwidth
if img.width > 512 or img.height > 512:
# Use higher-quality resampling for downsampling (LANCZOS)
img.thumbnail((512, 512), Image.Resampling.LANCZOS)
images.append(img)
return images
def taesd_preview(x: torch.Tensor, flux: bool = False, step: int = 0, total_steps: int = 0):
"""#### Preview the batched latent tensors as images.
#### Args:
- `x` (torch.Tensor): Input latent tensor with shape [B,C,H,W]
- `flux` (bool, optional): Whether using flux model (for channel ordering). Defaults to False.
- `step` (int): Current step number.
- `total_steps` (int): Total number of steps.
"""
if app_instance.app.previewer_var.get() is True:
try:
images = decode_latents_to_images(x, flux)
if images:
app_instance.app.update_image(images, step=step, total_steps=total_steps)
except Exception:
pass
|