File size: 14,865 Bytes
5b8131f | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 | """
MatFuse Condition Encoders for diffusers.
These encoders handle the multi-modal conditioning:
- Image embedding (CLIP image encoder)
- Text embedding (CLIP text encoder)
- Sketch encoder (CNN)
- Palette encoder (MLP)
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Optional, Dict, Union, List
from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.models.modeling_utils import ModelMixin
class SketchEncoder(ModelMixin, ConfigMixin):
"""
CNN encoder for binary sketch/edge maps.
Takes a single-channel binary image and encodes it to a spatial feature map
that will be concatenated with the latent for hybrid conditioning.
"""
@register_to_config
def __init__(
self,
in_channels: int = 1,
out_channels: int = 4,
):
super().__init__()
self.net = nn.Sequential(
nn.Conv2d(in_channels, 32, 7, 1, 1),
nn.BatchNorm2d(32),
nn.GELU(),
nn.Conv2d(32, 64, 3, 2, 1),
nn.BatchNorm2d(64),
nn.GELU(),
nn.Conv2d(64, 128, 3, 2, 1),
nn.BatchNorm2d(128),
nn.GELU(),
nn.Conv2d(128, 256, 3, 2, 1),
nn.BatchNorm2d(256),
nn.GELU(),
nn.Conv2d(256, out_channels, 1, 1, 0),
nn.BatchNorm2d(out_channels),
nn.GELU(),
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Encode sketch input.
Args:
x: Input tensor of shape (B, 1, H, W) with values in [0, 1].
Returns:
Encoded features of shape (B, out_channels, H/8, W/8).
"""
return self.net(x)
class PaletteEncoder(ModelMixin, ConfigMixin):
"""
MLP encoder for color palettes.
Takes a color palette (N colors, RGB) and encodes it to a single embedding
for cross-attention conditioning.
"""
@register_to_config
def __init__(
self,
in_channels: int = 3,
hidden_channels: int = 64,
out_channels: int = 512,
n_colors: int = 5,
):
super().__init__()
self.net = nn.Sequential(
nn.Linear(in_channels, hidden_channels),
nn.GELU(),
nn.Flatten(),
nn.Linear(hidden_channels * n_colors, out_channels),
nn.GELU(),
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Encode color palette.
Args:
x: Input tensor of shape (B, n_colors, 3) with RGB values in [0, 1].
Returns:
Encoded embedding of shape (B, out_channels).
"""
return self.net(x)
class CLIPImageEncoder(ModelMixin, ConfigMixin):
"""
Wrapper for CLIP image encoder using the OpenAI CLIP library.
Generates image embeddings for cross-attention conditioning.
"""
@register_to_config
def __init__(
self,
model_name: str = "ViT-B/16",
normalize: bool = True,
):
super().__init__()
self.model_name = model_name
self.normalize = normalize
self.model = None # Lazy loading
# Register normalization buffers
self.register_buffer(
"mean", torch.tensor([0.48145466, 0.4578275, 0.40821073]), persistent=False
)
self.register_buffer(
"std", torch.tensor([0.26862954, 0.26130258, 0.27577711]), persistent=False
)
def _load_model(self):
"""Lazy load the CLIP model."""
if self.model is None:
import clip
self.model, _ = clip.load(self.model_name, device="cpu", jit=False)
self.model = self.model.visual
def preprocess(self, x: torch.Tensor) -> torch.Tensor:
"""Preprocess images for CLIP."""
# Resize to 224x224
x = F.interpolate(
x, size=(224, 224), mode="bicubic", align_corners=True, antialias=True
)
# Normalize from [-1, 1] to [0, 1]
x = (x + 1.0) / 2.0
# Normalize according to CLIP - move mean/std to device if needed
mean = self.mean.to(x.device).view(1, 3, 1, 1)
std = self.std.to(x.device).view(1, 3, 1, 1)
x = (x - mean) / std
return x
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""
Encode image using CLIP.
Args:
x: Input tensor of shape (B, 3, H, W) with values in [-1, 1].
Returns:
Image embedding of shape (B, 1, 512).
"""
self._load_model()
# Move model to same device as input
device = x.device
self.model = self.model.to(device)
x = self.preprocess(x)
z = self.model(x).float().unsqueeze(1) # (B, 1, 512)
if self.normalize:
z = z / torch.linalg.norm(z, dim=2, keepdim=True)
return z
class CLIPTextEncoder(ModelMixin, ConfigMixin):
"""
Wrapper for CLIP sentence encoder using sentence-transformers.
Generates text embeddings for cross-attention conditioning.
"""
@register_to_config
def __init__(
self,
model_name: str = "sentence-transformers/clip-ViT-B-16",
):
super().__init__()
self.model_name = model_name
self.model = None # Lazy loading
def _load_model(self):
"""Lazy load the sentence transformer model."""
if self.model is None:
from sentence_transformers import SentenceTransformer
self.model = SentenceTransformer(self.model_name)
self.model.eval()
def forward(self, text: Union[str, List[str]]) -> torch.Tensor:
"""
Encode text using CLIP sentence transformer.
Args:
text: Input text or list of texts.
Returns:
Text embedding of shape (B, 512).
"""
self._load_model()
if isinstance(text, str):
text = [text]
embeddings = self.model.encode(text, convert_to_tensor=True)
return embeddings
class MultiConditionEncoder(ModelMixin, ConfigMixin):
"""
Multi-condition encoder that combines all conditioning modalities.
This encoder takes multiple condition inputs and produces:
- c_crossattn: Features for cross-attention (image, text, palette embeddings)
- c_concat: Features for concatenation (sketch encoding)
"""
@register_to_config
def __init__(
self,
sketch_in_channels: int = 1,
sketch_out_channels: int = 4,
palette_in_channels: int = 3,
palette_hidden_channels: int = 64,
palette_out_channels: int = 512,
n_colors: int = 5,
clip_image_model: str = "ViT-B/16",
clip_text_model: str = "sentence-transformers/clip-ViT-B-16",
):
super().__init__()
self.sketch_encoder = SketchEncoder(
in_channels=sketch_in_channels,
out_channels=sketch_out_channels,
)
self.palette_encoder = PaletteEncoder(
in_channels=palette_in_channels,
hidden_channels=palette_hidden_channels,
out_channels=palette_out_channels,
n_colors=n_colors,
)
# CLIP encoders are lazy-loaded
self.clip_image_encoder = None
self.clip_text_encoder = None
self._clip_image_model = clip_image_model
self._clip_text_model = clip_text_model
def _load_clip_encoders(self):
"""Lazy load CLIP encoders."""
if self.clip_image_encoder is None:
self.clip_image_encoder = CLIPImageEncoder(
model_name=self._clip_image_model
)
if self.clip_text_encoder is None:
self.clip_text_encoder = CLIPTextEncoder(model_name=self._clip_text_model)
def encode_image(self, image: torch.Tensor) -> torch.Tensor:
"""Encode image using CLIP."""
self._load_clip_encoders()
return self.clip_image_encoder(image)
def encode_text(self, text: Union[str, List[str]]) -> torch.Tensor:
"""Encode text using CLIP."""
self._load_clip_encoders()
return self.clip_text_encoder(text)
def encode_sketch(self, sketch: torch.Tensor) -> torch.Tensor:
"""Encode sketch/edge map."""
return self.sketch_encoder(sketch)
def encode_palette(self, palette: torch.Tensor) -> torch.Tensor:
"""Encode color palette."""
return self.palette_encoder(palette)
def get_unconditional_conditioning(
self,
batch_size: int = 1,
image_size: int = 256,
device: Optional[torch.device] = None,
) -> Dict[str, torch.Tensor]:
"""
Get unconditional conditioning for classifier-free guidance.
IMPORTANT: The original model was trained to drop conditions by replacing them
with encoded placeholders (zero/gray image through CLIP, empty string through
sentence-transformers, zero palette through PaletteEncoder, zero sketch through
SketchEncoder) — NOT with zero tensors. This method produces the correct
unconditional embeddings.
Args:
batch_size: Batch size.
image_size: Image resolution (for sketch spatial dims).
device: Device to place tensors on.
Returns:
Dictionary with c_crossattn and c_concat for unconditional guidance.
"""
return self.forward(
image_embed=None,
text=None,
sketch=None,
palette=None,
batch_size=batch_size,
image_size=image_size,
device=device,
)
def forward(
self,
image_embed: Optional[torch.Tensor] = None,
text: Optional[Union[str, List[str]]] = None,
sketch: Optional[torch.Tensor] = None,
palette: Optional[torch.Tensor] = None,
batch_size: int = 1,
image_size: int = 256,
device: Optional[torch.device] = None,
) -> Dict[str, torch.Tensor]:
"""
Encode all conditions.
When a condition is not provided, the model encodes a placeholder input
through the actual encoder (matching training behavior) rather than using
zero tensors. This is critical because the model was trained with:
- Image drop → CLIP encoding of a gray/zero image (0.0 in [-1,1])
- Text drop → sentence-transformer encoding of ""
- Palette drop → PaletteEncoder(zeros)
- Sketch drop → SketchEncoder(zeros)
Args:
image_embed: Reference image of shape (B, 3, H, W) in [-1, 1].
text: Text description(s).
sketch: Binary sketch of shape (B, 1, H, W) in [0, 1].
palette: Color palette of shape (B, n_colors, 3) in [0, 1].
batch_size: Batch size (used when no inputs are provided).
image_size: Image resolution (used to create placeholder sketch).
device: Device to place tensors on.
Returns:
Dictionary with:
- c_crossattn: Cross-attention context of shape (B, 3, 512) - always 3 tokens.
- c_concat: Concatenation features of shape (B, 4, H/8, W/8).
"""
self._load_clip_encoders()
# Determine batch size and device from any available input
if image_embed is not None:
batch_size = image_embed.shape[0]
device = device or image_embed.device
image_size = image_embed.shape[-1]
elif sketch is not None:
batch_size = sketch.shape[0]
device = device or sketch.device
image_size = sketch.shape[-1]
elif palette is not None:
batch_size = palette.shape[0]
device = device or palette.device
device = device or torch.device("cpu")
# Infer dtype from model weights for placeholder tensors (e.g. float16)
dtype = next(self.palette_encoder.parameters()).dtype
# --- Image embedding (token 0) ---
# When not provided, encode a zero (gray) image through CLIP, matching training ucg_training val=0.0
if image_embed is not None:
img_emb = self.clip_image_encoder(image_embed) # (B, 1, 512)
else:
placeholder_img = torch.zeros(
batch_size, 3, image_size, image_size, device=device, dtype=dtype
)
img_emb = self.clip_image_encoder(placeholder_img) # (B, 1, 512)
# --- Text embedding (token 1) ---
# When not provided, encode empty string through sentence-transformers, matching training ucg_training val=""
if text is not None:
text_emb = self.clip_text_encoder(text) # (B, 512)
if device is not None:
text_emb = text_emb.to(device)
text_emb = text_emb.unsqueeze(1) # (B, 1, 512)
else:
text_emb = self.clip_text_encoder([""] * batch_size) # (B, 512)
text_emb = text_emb.to(device).unsqueeze(1) # (B, 1, 512)
# --- Palette embedding (token 2) ---
# When not provided, encode zero palette through PaletteEncoder, matching training ucg_training val=0.0
if palette is not None:
palette_emb = self.palette_encoder(palette) # (B, 512)
palette_emb = palette_emb.unsqueeze(1) # (B, 1, 512)
else:
n_colors = self.config.get("n_colors", 5)
placeholder_palette = torch.zeros(batch_size, n_colors, 3, device=device, dtype=dtype)
palette_emb = self.palette_encoder(placeholder_palette) # (B, 512)
palette_emb = palette_emb.unsqueeze(1) # (B, 1, 512)
# Combine cross-attention embeddings - always (B, 3, 512)
c_crossattn = torch.cat([img_emb, text_emb, palette_emb], dim=1)
# --- Sketch encoding for concatenation ---
# When not provided, encode zero sketch through SketchEncoder, matching training ucg_training val=0.0
if sketch is not None:
c_concat = self.sketch_encoder(sketch) # (B, 4, H/8, W/8)
else:
placeholder_sketch = torch.zeros(
batch_size, 1, image_size, image_size, device=device, dtype=dtype
)
c_concat = self.sketch_encoder(placeholder_sketch) # (B, 4, H/8, W/8)
return {
"c_crossattn": c_crossattn,
"c_concat": c_concat,
}
|