Spaces:
Running
on
A100
Running
on
A100
File size: 17,200 Bytes
efa1353 8020398 efa1353 0639262 c812124 669715f 7d299b1 669715f bd18e87 669715f 39564a4 669715f 52cb438 669715f 0d04150 024ee6a 0d04150 024ee6a 0d04150 7983b33 669715f 0d04150 8cbd2c7 cb4eb0e 0d04150 c24728c 8020398 b6f23ea 0d04150 8020398 0d04150 52cb438 b6f23ea 52cb438 0d04150 52cb438 8020398 bd18e87 52cb438 bd18e87 cb4eb0e bd18e87 cb4eb0e 8020398 52cb438 b6f23ea 52cb438 8020398 0d04150 c718918 b6f23ea 8020398 6ec8160 8020398 52cb438 cb4eb0e bd18e87 cb4eb0e bd18e87 8020398 0d04150 8020398 669715f 32c9458 6b8527f 9d75037 32c9458 669715f 8c17f89 c718918 cdc3395 c718918 32c9458 c718918 32c9458 c718918 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 |
import sys
import os
from pathlib import Path
import gc
# Add the StableCascade and CSD directories to the Python path
app_dir = Path(__file__).parent
sys.path.extend([
str(app_dir),
str(app_dir / "third_party" / "StableCascade"),
str(app_dir / "third_party" / "CSD")
])
import yaml
import torch
from tqdm import tqdm
from accelerate.utils import set_module_tensor_to_device
import torch.nn.functional as F
import torchvision.transforms as T
from lang_sam import LangSAM
from inference.utils import *
from core.utils import load_or_fail
from train import WurstCoreC, WurstCoreB
from gdf_rbm import RBM
from stage_c_rbm import StageCRBM
from utils import WurstCoreCRBM
from gdf.schedulers import CosineSchedule
from gdf import VPScaler, CosineTNoiseCond, DDPMSampler, P2LossWeight, AdaptiveLossWeight
from gdf.targets import EpsilonTarget
import PIL
# Device configuration
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print(device)
# Flag for low VRAM usage
low_vram = True
# Function definition for low VRAM usage
def models_to(model, device="cpu", excepts=None):
"""
Change the device of nn.Modules within a class, skipping specified attributes.
"""
for attr_name in dir(model):
if attr_name.startswith('__') and attr_name.endswith('__'):
continue # skip special attributes
attr_value = getattr(model, attr_name, None)
if isinstance(attr_value, torch.nn.Module):
if excepts and attr_name in excepts:
print(f"Except '{attr_name}'")
continue
print(f"Change device of '{attr_name}' to {device}")
attr_value.to(device)
torch.cuda.empty_cache()
# Stage C model configuration
config_file = 'third_party/StableCascade/configs/inference/stage_c_3b.yaml'
with open(config_file, "r", encoding="utf-8") as file:
loaded_config = yaml.safe_load(file)
core = WurstCoreCRBM(config_dict=loaded_config, device=device, training=False)
# Stage B model configuration
config_file_b = 'third_party/StableCascade/configs/inference/stage_b_3b.yaml'
with open(config_file_b, "r", encoding="utf-8") as file:
config_file_b = yaml.safe_load(file)
core_b = WurstCoreB(config_dict=config_file_b, device=device, training=False)
# Setup extras and models for Stage C
extras = core.setup_extras_pre()
gdf_rbm = RBM(
schedule=CosineSchedule(clamp_range=[0.0001, 0.9999]),
input_scaler=VPScaler(), target=EpsilonTarget(),
noise_cond=CosineTNoiseCond(),
loss_weight=AdaptiveLossWeight(),
)
sampling_configs = {
"cfg": 5,
"sampler": DDPMSampler(gdf_rbm),
"shift": 1,
"timesteps": 20
}
extras = core.Extras(
gdf=gdf_rbm,
sampling_configs=sampling_configs,
transforms=extras.transforms,
effnet_preprocess=extras.effnet_preprocess,
clip_preprocess=extras.clip_preprocess
)
models = core.setup_models(extras)
models.generator.eval().requires_grad_(False)
# Setup extras and models for Stage B
extras_b = core_b.setup_extras_pre()
models_b = core_b.setup_models(extras_b, skip_clip=True)
models_b = WurstCoreB.Models(
**{**models_b.to_dict(), 'tokenizer': models.tokenizer, 'text_model': models.text_model}
)
models_b.generator.bfloat16().eval().requires_grad_(False)
# Off-load old generator (low VRAM mode)
if low_vram:
models.generator.to("cpu")
torch.cuda.empty_cache()
# Load and configure new generator
generator_rbm = StageCRBM()
for param_name, param in load_or_fail(core.config.generator_checkpoint_path).items():
set_module_tensor_to_device(generator_rbm, param_name, "cpu", value=param)
generator_rbm = generator_rbm.to(getattr(torch, core.config.dtype)).to(device)
generator_rbm = core.load_model(generator_rbm, 'generator')
# Create models_rbm instance
models_rbm = core.Models(
effnet=models.effnet,
text_model=models.text_model,
tokenizer=models.tokenizer,
generator=generator_rbm,
previewer=models.previewer,
image_model=models.image_model # Add this line
)
def reset_inference_state():
global models_rbm, models_b, extras, extras_b, device, core, core_b
# Reset sampling configurations
extras.sampling_configs['cfg'] = 5
extras.sampling_configs['shift'] = 1
extras.sampling_configs['timesteps'] = 20
extras.sampling_configs['t_start'] = 1.0
extras_b.sampling_configs['cfg'] = 1.1
extras_b.sampling_configs['shift'] = 1
extras_b.sampling_configs['timesteps'] = 10
extras_b.sampling_configs['t_start'] = 1.0
# Move models to CPU to free up GPU memory
models_to(models_rbm, device="cpu")
models_b.generator.to("cpu")
# Clear CUDA cache
torch.cuda.empty_cache()
gc.collect()
# Move necessary models back to the correct device
if low_vram:
models_to(models_rbm, device="cpu", excepts=["generator", "previewer"])
models_rbm.generator.to(device)
models_rbm.previewer.to(device)
else:
models_to(models_rbm, device=device)
models_b.generator.to("cpu") # Keep Stage B generator on CPU for now
# Ensure effnet and image_model are on the correct device
models_rbm.effnet.to(device)
if models_rbm.image_model is not None:
models_rbm.image_model.to(device)
# Reset model states
models_rbm.generator.eval().requires_grad_(False)
models_b.generator.bfloat16().eval().requires_grad_(False)
# Clear CUDA cache again
torch.cuda.empty_cache()
gc.collect()
def infer(ref_style_file, style_description, caption):
global models_rbm, models_b
try:
height=1024
width=1024
batch_size=1
output_file='output.png'
stage_c_latent_shape, stage_b_latent_shape = calculate_latent_sizes(height, width, batch_size=batch_size)
extras.sampling_configs['cfg'] = 4
extras.sampling_configs['shift'] = 2
extras.sampling_configs['timesteps'] = 20
extras.sampling_configs['t_start'] = 1.0
extras_b.sampling_configs['cfg'] = 1.1
extras_b.sampling_configs['shift'] = 1
extras_b.sampling_configs['timesteps'] = 10
extras_b.sampling_configs['t_start'] = 1.0
ref_style = resize_image(PIL.Image.open(ref_style_file).convert("RGB")).unsqueeze(0).expand(batch_size, -1, -1, -1).to(device)
batch = {'captions': [caption] * batch_size}
batch['style'] = ref_style
# Ensure models are on the correct device before inference
if low_vram:
models_to(models_rbm, device=device, excepts=["generator", "previewer"])
else:
models_to(models_rbm, device=device)
models_b.generator.to(device)
# Ensure effnet and image_model are on the correct device
models_rbm.effnet.to(device)
if models_rbm.image_model is not None:
models_rbm.image_model.to(device)
x0_style_forward = models_rbm.effnet(extras.effnet_preprocess(ref_style))
conditions = core.get_conditions(batch, models_rbm, extras, is_eval=True, is_unconditional=False, eval_image_embeds=True, eval_style=True, eval_csd=False)
unconditions = core.get_conditions(batch, models_rbm, extras, is_eval=True, is_unconditional=True, eval_image_embeds=False)
conditions_b = core_b.get_conditions(batch, models_b, extras_b, is_eval=True, is_unconditional=False)
unconditions_b = core_b.get_conditions(batch, models_b, extras_b, is_eval=True, is_unconditional=True)
if low_vram:
# The sampling process uses more vram, so we offload everything except two modules to the cpu.
models_to(models_rbm, device="cpu", excepts=["generator", "previewer"])
# Stage C reverse process.
sampling_c = extras.gdf.sample(
models_rbm.generator, conditions, stage_c_latent_shape,
unconditions, device=device,
**extras.sampling_configs,
x0_style_forward=x0_style_forward,
apply_pushforward=False, tau_pushforward=8,
num_iter=3, eta=0.1, tau=20, eval_csd=True,
extras=extras, models=models_rbm,
lam_style=1, lam_txt_alignment=1.0,
use_ddim_sampler=True,
)
for (sampled_c, _, _) in tqdm(sampling_c, total=extras.sampling_configs['timesteps']):
sampled_c = sampled_c
# Stage B reverse process.
with torch.no_grad(), torch.cuda.amp.autocast(dtype=torch.bfloat16):
conditions_b['effnet'] = sampled_c
unconditions_b['effnet'] = torch.zeros_like(sampled_c)
sampling_b = extras_b.gdf.sample(
models_b.generator, conditions_b, stage_b_latent_shape,
unconditions_b, device=device, **extras_b.sampling_configs,
)
for (sampled_b, _, _) in tqdm(sampling_b, total=extras_b.sampling_configs['timesteps']):
sampled_b = sampled_b
sampled = models_b.stage_a.decode(sampled_b).float()
sampled = torch.cat([
torch.nn.functional.interpolate(ref_style.cpu(), size=(height, width)),
sampled.cpu(),
], dim=0)
# Remove the batch dimension and keep only the generated image
sampled = sampled[1] # This selects the generated image, discarding the reference style image
# Ensure the tensor is in [C, H, W] format
if sampled.dim() == 3 and sampled.shape[0] == 3:
sampled_image = T.ToPILImage()(sampled) # Convert tensor to PIL image
sampled_image.save(output_file) # Save the image as a PNG
else:
raise ValueError(f"Expected tensor of shape [3, H, W] but got {sampled.shape}")
return output_file # Return the path to the saved image
finally:
# Reset the state after inference, regardless of success or failure
reset_inference_state()
def infer_compo(style_description, ref_style_file, caption, ref_sub_file):
global models_rbm, models_b
try:
caption = f"{caption} in {style_description}"
sam_prompt = f"{caption}"
use_sam_mask = False
if low_vram:
# Revert the devices of the modules back to their original state
models_to(models_rbm, device)
batch_size = 1
height, width = 1024, 1024
stage_c_latent_shape, stage_b_latent_shape = calculate_latent_sizes(height, width, batch_size=batch_size)
extras.sampling_configs['cfg'] = 4
extras.sampling_configs['shift'] = 2
extras.sampling_configs['timesteps'] = 20
extras.sampling_configs['t_start'] = 1.0
extras_b.sampling_configs['cfg'] = 1.1
extras_b.sampling_configs['shift'] = 1
extras_b.sampling_configs['timesteps'] = 10
extras_b.sampling_configs['t_start'] = 1.0
ref_style = resize_image(PIL.Image.open(ref_style_file).convert("RGB")).unsqueeze(0).expand(batch_size, -1, -1, -1).to(device)
ref_images = resize_image(PIL.Image.open(ref_sub_file).convert("RGB")).unsqueeze(0).expand(batch_size, -1, -1, -1).to(device)
batch = {'captions': [caption] * batch_size}
batch['style'] = ref_style
batch['images'] = ref_images
x0_forward = models_rbm.effnet(extras.effnet_preprocess(ref_images.to(device)))
x0_style_forward = models_rbm.effnet(extras.effnet_preprocess(ref_style.to(device)))
## SAM Mask for sub
use_sam_mask = False
x0_preview = models_rbm.previewer(x0_forward)
sam_model = LangSAM()
# Convert tensor to PIL Image before passing to sam_model.predict
x0_preview_pil = T.ToPILImage()(x0_preview[0])
x0_preview_tensor = T.ToTensor()(x0_preview_pil) # Convert PIL Image back to tensor
sam_mask, boxes, phrases, logits = sam_model.predict(x0_preview_pil, sam_prompt)
sam_mask = sam_mask.detach().unsqueeze(dim=0).to(device)
conditions = core.get_conditions(batch, models_rbm, extras, is_eval=True, is_unconditional=False, eval_image_embeds=True, eval_subject_style=True, eval_csd=False)
unconditions = core.get_conditions(batch, models_rbm, extras, is_eval=True, is_unconditional=True, eval_image_embeds=False, eval_subject_style=True)
conditions_b = core_b.get_conditions(batch, models_b, extras_b, is_eval=True, is_unconditional=False)
unconditions_b = core_b.get_conditions(batch, models_b, extras_b, is_eval=True, is_unconditional=True)
if low_vram:
# The sampling process uses more vram, so we offload everything except two modules to the cpu.
models_to(models_rbm, device="cpu", excepts=["generator", "previewer"])
models_to(sam_model, device="cpu")
models_to(sam_model.sam, device="cpu")
# Stage C reverse process.
sampling_c = extras.gdf.sample(
models_rbm.generator, conditions, stage_c_latent_shape,
unconditions, device=device,
**extras.sampling_configs,
x0_style_forward=x0_style_forward, x0_forward=x0_forward,
apply_pushforward=False, tau_pushforward=5, tau_pushforward_csd=10,
num_iter=3, eta=1e-1, tau=20, eval_sub_csd=True,
extras=extras, models=models_rbm,
use_attn_mask=use_sam_mask,
save_attn_mask=False,
lam_content=1, lam_style=1,
sam_mask=sam_mask, use_sam_mask=use_sam_mask,
sam_prompt=sam_prompt
)
for (sampled_c, _, _) in tqdm(sampling_c, total=extras.sampling_configs['timesteps']):
sampled_c = sampled_c
# Stage B reverse process.
with torch.no_grad(), torch.cuda.amp.autocast(dtype=torch.bfloat16):
conditions_b['effnet'] = sampled_c
unconditions_b['effnet'] = torch.zeros_like(sampled_c)
sampling_b = extras_b.gdf.sample(
models_b.generator, conditions_b, stage_b_latent_shape,
unconditions_b, device=device, **extras_b.sampling_configs,
)
for (sampled_b, _, _) in tqdm(sampling_b, total=extras_b.sampling_configs['timesteps']):
sampled_b = sampled_b
sampled = models_b.stage_a.decode(sampled_b).float()
sampled = torch.cat([
torch.nn.functional.interpolate(ref_images.cpu(), size=(height, width)),
torch.nn.functional.interpolate(ref_style.cpu(), size=(height, width)),
sampled.cpu(),
], dim=0)
# Remove the batch dimension and keep only the generated image
sampled = sampled[2] # This selects the generated image, discarding the reference images
# Ensure the tensor is in [C, H, W] format
if sampled.dim() == 3 and sampled.shape[0] == 3:
output_file = 'output_compo.png'
sampled_image = T.ToPILImage()(sampled) # Convert tensor to PIL image
sampled_image.save(output_file) # Save the image as a PNG
else:
raise ValueError(f"Expected tensor of shape [3, H, W] but got {sampled.shape}")
return output_file # Return the path to the saved image
finally:
# Reset the state after inference, regardless of success or failure
reset_inference_state()
import gradio as gr
with gr.Blocks() as demo:
with gr.Column():
gr.Markdown("# RB-Modulation")
gr.Markdown("## Training-Free Personalization of Diffusion Models using Stochastic Optimal Control")
gr.HTML("""
<div style="display:flex;column-gap:4px;">
<a href='https://rb-modulation.github.io'>
<img src='https://img.shields.io/badge/Project-Page-Green'>
</a>
<a href='https://arxiv.org/pdf/2405.17401'>
<img src='https://img.shields.io/badge/Paper-Arxiv-red'>
</a>
</div>
""")
with gr.Row():
with gr.Column():
style_reference_image = gr.Image(
label = "Style Reference Image",
type = "filepath"
)
style_description = gr.Textbox(
label ="Style Description"
)
subject_prompt = gr.Textbox(
label = "Subject Prompt"
)
with gr.Accordion("Advanced Settings", open=False):
subject_reference = gr.Image(type="filepath")
use_subject_ref = gr.Checkbox(label="Use Subject Image as Reference", value=False)
submit_btn = gr.Button("Submit")
with gr.Column():
output_image = gr.Image(label="Output Image")
'''
submit_btn.click(
fn = infer,
inputs = [style_reference_image, style_description, subject_prompt],
outputs = [output_image]
)
'''
submit_btn.click(
fn = infer_compo,
inputs = [style_description, style_reference_image, subject_prompt, subject_reference],
outputs = [output_image]
)
demo.launch() |