Spaces:
Running
on
Zero
Running
on
Zero
zhengchong
commited on
Commit
•
905c952
1
Parent(s):
3e791eb
feat: Enhance CatVTON functionality with new pipelines and UI improvements
Browse files- Added CatVTONPix2PixPipeline and FluxTryOnPipeline to support additional virtual try-on methods.
- Implemented new submit functions for mask-free and Flux-based try-on.
- Updated UI to include separate tabs for mask-based and mask-free options, enhancing user experience.
- Modified requirements.txt to include new dependencies and updated existing ones.
- Improved error handling and image processing in the submission functions.
- app.py +451 -113
- model/flux/pipeline_flux_tryon.py +499 -0
- model/flux/transformer_flux.py +672 -0
- model/pipeline.py +117 -0
- requirements.txt +4 -3
app.py
CHANGED
@@ -13,7 +13,8 @@ from huggingface_hub import snapshot_download
|
|
13 |
from PIL import Image
|
14 |
torch.jit.script = lambda f: f
|
15 |
from model.cloth_masker import AutoMasker, vis_mask
|
16 |
-
from model.pipeline import CatVTONPipeline
|
|
|
17 |
from utils import init_weight_dtype, resize_and_crop, resize_and_padding
|
18 |
|
19 |
|
@@ -105,7 +106,10 @@ def image_grid(imgs, rows, cols):
|
|
105 |
|
106 |
|
107 |
args = parse_args()
|
108 |
-
|
|
|
|
|
|
|
109 |
# Pipeline
|
110 |
pipeline = CatVTONPipeline(
|
111 |
base_ckpt=args.base_model_path,
|
@@ -123,6 +127,30 @@ automasker = AutoMasker(
|
|
123 |
device='cuda',
|
124 |
)
|
125 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
126 |
@spaces.GPU(duration=120)
|
127 |
def submit_function(
|
128 |
person_image,
|
@@ -202,10 +230,135 @@ def submit_function(
|
|
202 |
new_result_image.paste(result_image, (condition_width + 5, 0))
|
203 |
return new_result_image
|
204 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
205 |
|
206 |
def person_example_fn(image_path):
|
207 |
return image_path
|
208 |
|
|
|
209 |
HEADER = """
|
210 |
<h1 style="text-align: center;"> 🐈 CatVTON: Concatenation Is All You Need for Virtual Try-On with Diffusion Models </h1>
|
211 |
<div style="display: flex; justify-content: center; align-items: center;">
|
@@ -241,136 +394,321 @@ HEADER = """
|
|
241 |
def app_gradio():
|
242 |
with gr.Blocks(title="CatVTON") as demo:
|
243 |
gr.Markdown(HEADER)
|
244 |
-
with gr.
|
245 |
-
with gr.
|
246 |
-
with gr.
|
247 |
-
|
248 |
-
|
249 |
-
|
250 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
251 |
)
|
252 |
-
|
253 |
-
|
|
|
254 |
)
|
255 |
-
|
256 |
-
|
257 |
-
|
258 |
-
|
259 |
-
|
|
|
|
|
260 |
)
|
261 |
-
|
262 |
-
gr.
|
263 |
-
|
264 |
)
|
265 |
-
|
266 |
-
label="
|
267 |
-
choices=["
|
268 |
-
value="
|
269 |
)
|
270 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
271 |
|
272 |
-
|
273 |
-
|
274 |
-
'<center><span style="color: #FF0000">!!! Click only Once, Wait for Delay !!!</span></center>'
|
275 |
)
|
276 |
-
|
277 |
-
|
278 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
279 |
)
|
280 |
-
|
281 |
-
|
282 |
-
|
283 |
-
|
284 |
-
|
285 |
-
|
286 |
-
|
287 |
-
|
288 |
-
|
289 |
-
|
290 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
291 |
)
|
292 |
-
|
293 |
-
|
294 |
-
|
295 |
-
value="input & mask & result",
|
296 |
)
|
297 |
-
|
298 |
-
|
299 |
-
|
300 |
-
|
301 |
-
|
302 |
-
|
303 |
-
|
304 |
-
|
305 |
-
|
306 |
-
|
307 |
-
|
308 |
-
],
|
309 |
-
examples_per_page=4,
|
310 |
-
inputs=image_path,
|
311 |
-
label="Person Examples ①",
|
312 |
)
|
313 |
-
|
314 |
-
|
315 |
-
|
316 |
-
|
317 |
-
|
318 |
-
|
319 |
-
|
320 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
321 |
)
|
322 |
-
gr.
|
323 |
-
|
324 |
)
|
325 |
-
|
326 |
-
|
327 |
-
|
328 |
-
|
329 |
-
|
330 |
-
|
331 |
-
|
332 |
-
|
333 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
334 |
)
|
335 |
-
|
336 |
-
|
337 |
-
|
338 |
-
for _ in os.listdir(os.path.join(root_path, "condition", "overall"))
|
339 |
-
],
|
340 |
-
examples_per_page=4,
|
341 |
-
inputs=cloth_image,
|
342 |
-
label="Condition Overall Examples",
|
343 |
)
|
344 |
-
|
345 |
-
|
346 |
-
|
347 |
-
for _ in os.listdir(os.path.join(root_path, "condition", "person"))
|
348 |
-
],
|
349 |
-
examples_per_page=4,
|
350 |
-
inputs=cloth_image,
|
351 |
-
label="Condition Reference Person Examples",
|
352 |
)
|
353 |
-
gr.
|
354 |
-
|
|
|
|
|
355 |
)
|
356 |
|
357 |
-
|
358 |
-
|
359 |
-
|
360 |
-
|
361 |
-
|
362 |
-
|
363 |
-
|
364 |
-
|
365 |
-
|
366 |
-
cloth_type,
|
367 |
-
|
368 |
-
|
369 |
-
|
370 |
-
show_type,
|
371 |
-
],
|
372 |
-
result_image,
|
373 |
-
)
|
374 |
demo.queue().launch(share=True, show_error=True)
|
375 |
|
376 |
|
|
|
13 |
from PIL import Image
|
14 |
torch.jit.script = lambda f: f
|
15 |
from model.cloth_masker import AutoMasker, vis_mask
|
16 |
+
from model.pipeline import CatVTONPipeline, CatVTONPix2PixPipeline
|
17 |
+
from model.flux.pipeline_flux_tryon import FluxTryOnPipeline
|
18 |
from utils import init_weight_dtype, resize_and_crop, resize_and_padding
|
19 |
|
20 |
|
|
|
106 |
|
107 |
|
108 |
args = parse_args()
|
109 |
+
|
110 |
+
# Mask-based CatVTON
|
111 |
+
catvton_repo = "zhengchong/CatVTON"
|
112 |
+
repo_path = snapshot_download(repo_id=catvton_repo)
|
113 |
# Pipeline
|
114 |
pipeline = CatVTONPipeline(
|
115 |
base_ckpt=args.base_model_path,
|
|
|
127 |
device='cuda',
|
128 |
)
|
129 |
|
130 |
+
|
131 |
+
# Flux-based CatVTON
|
132 |
+
flux_repo = "black-forest-labs/FLUX.1-Fill-dev"
|
133 |
+
pipeline_flux = FluxTryOnPipeline.from_pretrained(flux_repo)
|
134 |
+
pipeline_flux.load_lora_weights(
|
135 |
+
os.path.join(repo_path, "flux-lora"),
|
136 |
+
weight_name='pytorch_lora_weights.safetensors'
|
137 |
+
)
|
138 |
+
pipeline_flux.to("cuda", init_weight_dtype(args.mixed_precision))
|
139 |
+
|
140 |
+
|
141 |
+
# Mask-free CatVTON
|
142 |
+
catvton_mf_repo = "zhengchong/CatVTON-MaskFree"
|
143 |
+
repo_path_mf = snapshot_download(repo_id=catvton_mf_repo)
|
144 |
+
pipeline_p2p = CatVTONPix2PixPipeline(
|
145 |
+
base_ckpt=args.p2p_base_model_path,
|
146 |
+
attn_ckpt=repo_path,
|
147 |
+
attn_ckpt_version="mix-48k-1024",
|
148 |
+
weight_dtype=init_weight_dtype(args.mixed_precision),
|
149 |
+
use_tf32=args.allow_tf32,
|
150 |
+
device='cuda'
|
151 |
+
)
|
152 |
+
|
153 |
+
|
154 |
@spaces.GPU(duration=120)
|
155 |
def submit_function(
|
156 |
person_image,
|
|
|
230 |
new_result_image.paste(result_image, (condition_width + 5, 0))
|
231 |
return new_result_image
|
232 |
|
233 |
+
@spaces.GPU(duration=120)
|
234 |
+
def submit_function_p2p(
|
235 |
+
person_image,
|
236 |
+
cloth_image,
|
237 |
+
num_inference_steps,
|
238 |
+
guidance_scale,
|
239 |
+
seed):
|
240 |
+
person_image= person_image["background"]
|
241 |
+
|
242 |
+
tmp_folder = args.output_dir
|
243 |
+
date_str = datetime.now().strftime("%Y%m%d%H%M%S")
|
244 |
+
result_save_path = os.path.join(tmp_folder, date_str[:8], date_str[8:] + ".png")
|
245 |
+
if not os.path.exists(os.path.join(tmp_folder, date_str[:8])):
|
246 |
+
os.makedirs(os.path.join(tmp_folder, date_str[:8]))
|
247 |
+
|
248 |
+
generator = None
|
249 |
+
if seed != -1:
|
250 |
+
generator = torch.Generator(device='cuda').manual_seed(seed)
|
251 |
+
|
252 |
+
person_image = Image.open(person_image).convert("RGB")
|
253 |
+
cloth_image = Image.open(cloth_image).convert("RGB")
|
254 |
+
person_image = resize_and_crop(person_image, (args.width, args.height))
|
255 |
+
cloth_image = resize_and_padding(cloth_image, (args.width, args.height))
|
256 |
+
|
257 |
+
# Inference
|
258 |
+
try:
|
259 |
+
result_image = pipeline_p2p(
|
260 |
+
image=person_image,
|
261 |
+
condition_image=cloth_image,
|
262 |
+
num_inference_steps=num_inference_steps,
|
263 |
+
guidance_scale=guidance_scale,
|
264 |
+
generator=generator
|
265 |
+
)[0]
|
266 |
+
except Exception as e:
|
267 |
+
raise gr.Error(
|
268 |
+
"An error occurred. Please try again later: {}".format(e)
|
269 |
+
)
|
270 |
+
|
271 |
+
# Post-process
|
272 |
+
save_result_image = image_grid([person_image, cloth_image, result_image], 1, 3)
|
273 |
+
save_result_image.save(result_save_path)
|
274 |
+
return result_image
|
275 |
+
|
276 |
+
@spaces.GPU(duration=120)
|
277 |
+
def submit_function_flux(
|
278 |
+
person_image,
|
279 |
+
cloth_image,
|
280 |
+
cloth_type,
|
281 |
+
resolution,
|
282 |
+
num_inference_steps,
|
283 |
+
guidance_scale,
|
284 |
+
seed,
|
285 |
+
show_type
|
286 |
+
):
|
287 |
+
# Set height and width based on resolution
|
288 |
+
height = resolution
|
289 |
+
width = int(height * 0.75)
|
290 |
+
args.width = width
|
291 |
+
args.height = height
|
292 |
+
|
293 |
+
# Process image editor input
|
294 |
+
person_image, mask = person_image["background"], person_image["layers"][0]
|
295 |
+
mask = Image.open(mask).convert("L")
|
296 |
+
if len(np.unique(np.array(mask))) == 1:
|
297 |
+
mask = None
|
298 |
+
else:
|
299 |
+
mask = np.array(mask)
|
300 |
+
mask[mask > 0] = 255
|
301 |
+
mask = Image.fromarray(mask)
|
302 |
+
|
303 |
+
# Set random seed
|
304 |
+
generator = None
|
305 |
+
if seed != -1:
|
306 |
+
generator = torch.Generator(device='cuda').manual_seed(seed)
|
307 |
+
|
308 |
+
# Process input images
|
309 |
+
person_image = Image.open(person_image).convert("RGB")
|
310 |
+
cloth_image = Image.open(cloth_image).convert("RGB")
|
311 |
+
|
312 |
+
# Adjust image sizes
|
313 |
+
person_image = resize_and_crop(person_image, (args.width, args.height))
|
314 |
+
cloth_image = resize_and_padding(cloth_image, (args.width, args.height))
|
315 |
+
|
316 |
+
# Process mask
|
317 |
+
if mask is not None:
|
318 |
+
mask = resize_and_crop(mask, (args.width, args.height))
|
319 |
+
else:
|
320 |
+
mask = automasker(
|
321 |
+
person_image,
|
322 |
+
cloth_type
|
323 |
+
)['mask']
|
324 |
+
mask = mask_processor.blur(mask, blur_factor=9)
|
325 |
+
|
326 |
+
# Inference
|
327 |
+
result_image = pipeline_flux(
|
328 |
+
image=person_image,
|
329 |
+
condition_image=cloth_image,
|
330 |
+
mask=mask,
|
331 |
+
num_inference_steps=num_inference_steps,
|
332 |
+
guidance_scale=guidance_scale,
|
333 |
+
generator=generator
|
334 |
+
)[0]
|
335 |
+
|
336 |
+
# Post-processing
|
337 |
+
masked_person = vis_mask(person_image, mask)
|
338 |
+
|
339 |
+
# Return result based on show type
|
340 |
+
if show_type == "result only":
|
341 |
+
return result_image
|
342 |
+
else:
|
343 |
+
width, height = person_image.size
|
344 |
+
if show_type == "input & result":
|
345 |
+
condition_width = width // 2
|
346 |
+
conditions = image_grid([person_image, cloth_image], 2, 1)
|
347 |
+
else:
|
348 |
+
condition_width = width // 3
|
349 |
+
conditions = image_grid([person_image, masked_person, cloth_image], 3, 1)
|
350 |
+
|
351 |
+
conditions = conditions.resize((condition_width, height), Image.NEAREST)
|
352 |
+
new_result_image = Image.new("RGB", (width + condition_width + 5, height))
|
353 |
+
new_result_image.paste(conditions, (0, 0))
|
354 |
+
new_result_image.paste(result_image, (condition_width + 5, 0))
|
355 |
+
return new_result_image
|
356 |
+
|
357 |
|
358 |
def person_example_fn(image_path):
|
359 |
return image_path
|
360 |
|
361 |
+
|
362 |
HEADER = """
|
363 |
<h1 style="text-align: center;"> 🐈 CatVTON: Concatenation Is All You Need for Virtual Try-On with Diffusion Models </h1>
|
364 |
<div style="display: flex; justify-content: center; align-items: center;">
|
|
|
394 |
def app_gradio():
|
395 |
with gr.Blocks(title="CatVTON") as demo:
|
396 |
gr.Markdown(HEADER)
|
397 |
+
with gr.Tab("Mask-based & SD1.5"):
|
398 |
+
with gr.Row():
|
399 |
+
with gr.Column(scale=1, min_width=350):
|
400 |
+
with gr.Row():
|
401 |
+
image_path = gr.Image(
|
402 |
+
type="filepath",
|
403 |
+
interactive=True,
|
404 |
+
visible=False,
|
405 |
+
)
|
406 |
+
person_image = gr.ImageEditor(
|
407 |
+
interactive=True, label="Person Image", type="filepath"
|
408 |
+
)
|
409 |
+
|
410 |
+
with gr.Row():
|
411 |
+
with gr.Column(scale=1, min_width=230):
|
412 |
+
cloth_image = gr.Image(
|
413 |
+
interactive=True, label="Condition Image", type="filepath"
|
414 |
+
)
|
415 |
+
with gr.Column(scale=1, min_width=120):
|
416 |
+
gr.Markdown(
|
417 |
+
'<span style="color: #808080; font-size: small;">Two ways to provide Mask:<br>1. Upload the person image and use the `🖌️` above to draw the Mask (higher priority)<br>2. Select the `Try-On Cloth Type` to generate automatically </span>'
|
418 |
+
)
|
419 |
+
cloth_type = gr.Radio(
|
420 |
+
label="Try-On Cloth Type",
|
421 |
+
choices=["upper", "lower", "overall"],
|
422 |
+
value="upper",
|
423 |
+
)
|
424 |
+
|
425 |
+
|
426 |
+
submit = gr.Button("Submit")
|
427 |
+
gr.Markdown(
|
428 |
+
'<center><span style="color: #FF0000">!!! Click only Once, Wait for Delay !!!</span></center>'
|
429 |
)
|
430 |
+
|
431 |
+
gr.Markdown(
|
432 |
+
'<span style="color: #808080; font-size: small;">Advanced options can adjust details:<br>1. `Inference Step` may enhance details;<br>2. `CFG` is highly correlated with saturation;<br>3. `Random seed` may improve pseudo-shadow.</span>'
|
433 |
)
|
434 |
+
with gr.Accordion("Advanced Options", open=False):
|
435 |
+
num_inference_steps = gr.Slider(
|
436 |
+
label="Inference Step", minimum=10, maximum=100, step=5, value=50
|
437 |
+
)
|
438 |
+
# Guidence Scale
|
439 |
+
guidance_scale = gr.Slider(
|
440 |
+
label="CFG Strenth", minimum=0.0, maximum=7.5, step=0.5, value=2.5
|
441 |
)
|
442 |
+
# Random Seed
|
443 |
+
seed = gr.Slider(
|
444 |
+
label="Seed", minimum=-1, maximum=10000, step=1, value=42
|
445 |
)
|
446 |
+
show_type = gr.Radio(
|
447 |
+
label="Show Type",
|
448 |
+
choices=["result only", "input & result", "input & mask & result"],
|
449 |
+
value="input & mask & result",
|
450 |
)
|
451 |
|
452 |
+
with gr.Column(scale=2, min_width=500):
|
453 |
+
result_image = gr.Image(interactive=False, label="Result")
|
454 |
+
with gr.Row():
|
455 |
+
# Photo Examples
|
456 |
+
root_path = "resource/demo/example"
|
457 |
+
with gr.Column():
|
458 |
+
men_exm = gr.Examples(
|
459 |
+
examples=[
|
460 |
+
os.path.join(root_path, "person", "men", _)
|
461 |
+
for _ in os.listdir(os.path.join(root_path, "person", "men"))
|
462 |
+
],
|
463 |
+
examples_per_page=4,
|
464 |
+
inputs=image_path,
|
465 |
+
label="Person Examples ①",
|
466 |
+
)
|
467 |
+
women_exm = gr.Examples(
|
468 |
+
examples=[
|
469 |
+
os.path.join(root_path, "person", "women", _)
|
470 |
+
for _ in os.listdir(os.path.join(root_path, "person", "women"))
|
471 |
+
],
|
472 |
+
examples_per_page=4,
|
473 |
+
inputs=image_path,
|
474 |
+
label="Person Examples ②",
|
475 |
+
)
|
476 |
+
gr.Markdown(
|
477 |
+
'<span style="color: #808080; font-size: small;">*Person examples come from the demos of <a href="https://huggingface.co/spaces/levihsu/OOTDiffusion">OOTDiffusion</a> and <a href="https://www.outfitanyone.org">OutfitAnyone</a>. </span>'
|
478 |
+
)
|
479 |
+
with gr.Column():
|
480 |
+
condition_upper_exm = gr.Examples(
|
481 |
+
examples=[
|
482 |
+
os.path.join(root_path, "condition", "upper", _)
|
483 |
+
for _ in os.listdir(os.path.join(root_path, "condition", "upper"))
|
484 |
+
],
|
485 |
+
examples_per_page=4,
|
486 |
+
inputs=cloth_image,
|
487 |
+
label="Condition Upper Examples",
|
488 |
+
)
|
489 |
+
condition_overall_exm = gr.Examples(
|
490 |
+
examples=[
|
491 |
+
os.path.join(root_path, "condition", "overall", _)
|
492 |
+
for _ in os.listdir(os.path.join(root_path, "condition", "overall"))
|
493 |
+
],
|
494 |
+
examples_per_page=4,
|
495 |
+
inputs=cloth_image,
|
496 |
+
label="Condition Overall Examples",
|
497 |
+
)
|
498 |
+
condition_person_exm = gr.Examples(
|
499 |
+
examples=[
|
500 |
+
os.path.join(root_path, "condition", "person", _)
|
501 |
+
for _ in os.listdir(os.path.join(root_path, "condition", "person"))
|
502 |
+
],
|
503 |
+
examples_per_page=4,
|
504 |
+
inputs=cloth_image,
|
505 |
+
label="Condition Reference Person Examples",
|
506 |
+
)
|
507 |
+
gr.Markdown(
|
508 |
+
'<span style="color: #808080; font-size: small;">*Condition examples come from the Internet. </span>'
|
509 |
+
)
|
510 |
|
511 |
+
image_path.change(
|
512 |
+
person_example_fn, inputs=image_path, outputs=person_image
|
|
|
513 |
)
|
514 |
+
|
515 |
+
submit.click(
|
516 |
+
submit_function,
|
517 |
+
[
|
518 |
+
person_image,
|
519 |
+
cloth_image,
|
520 |
+
cloth_type,
|
521 |
+
num_inference_steps,
|
522 |
+
guidance_scale,
|
523 |
+
seed,
|
524 |
+
show_type,
|
525 |
+
],
|
526 |
+
result_image,
|
527 |
)
|
528 |
+
|
529 |
+
with gr.Tab("Mask-free & SD1.5"):
|
530 |
+
with gr.Row():
|
531 |
+
with gr.Column(scale=1, min_width=350):
|
532 |
+
with gr.Row():
|
533 |
+
image_path_p2p = gr.Image(
|
534 |
+
type="filepath",
|
535 |
+
interactive=True,
|
536 |
+
visible=False,
|
537 |
+
)
|
538 |
+
person_image_p2p = gr.ImageEditor(
|
539 |
+
interactive=True, label="Person Image", type="filepath"
|
540 |
+
)
|
541 |
+
|
542 |
+
with gr.Row():
|
543 |
+
with gr.Column(scale=1, min_width=230):
|
544 |
+
cloth_image_p2p = gr.Image(
|
545 |
+
interactive=True, label="Condition Image", type="filepath"
|
546 |
+
)
|
547 |
+
|
548 |
+
submit_p2p = gr.Button("Submit")
|
549 |
+
gr.Markdown(
|
550 |
+
'<center><span style="color: #FF0000">!!! Click only Once, Wait for Delay !!!</span></center>'
|
551 |
)
|
552 |
+
|
553 |
+
gr.Markdown(
|
554 |
+
'<span style="color: #808080; font-size: small;">Advanced options can adjust details:<br>1. `Inference Step` may enhance details;<br>2. `CFG` is highly correlated with saturation;<br>3. `Random seed` may improve pseudo-shadow.</span>'
|
|
|
555 |
)
|
556 |
+
with gr.Accordion("Advanced Options", open=False):
|
557 |
+
num_inference_steps_p2p = gr.Slider(
|
558 |
+
label="Inference Step", minimum=10, maximum=100, step=5, value=50
|
559 |
+
)
|
560 |
+
# Guidence Scale
|
561 |
+
guidance_scale_p2p = gr.Slider(
|
562 |
+
label="CFG Strenth", minimum=0.0, maximum=7.5, step=0.5, value=2.5
|
563 |
+
)
|
564 |
+
# Random Seed
|
565 |
+
seed_p2p = gr.Slider(
|
566 |
+
label="Seed", minimum=-1, maximum=10000, step=1, value=42
|
|
|
|
|
|
|
|
|
567 |
)
|
568 |
+
# show_type = gr.Radio(
|
569 |
+
# label="Show Type",
|
570 |
+
# choices=["result only", "input & result", "input & mask & result"],
|
571 |
+
# value="input & mask & result",
|
572 |
+
# )
|
573 |
+
|
574 |
+
with gr.Column(scale=2, min_width=500):
|
575 |
+
result_image_p2p = gr.Image(interactive=False, label="Result")
|
576 |
+
with gr.Row():
|
577 |
+
# Photo Examples
|
578 |
+
root_path = "resource/demo/example"
|
579 |
+
with gr.Column():
|
580 |
+
gr.Examples(
|
581 |
+
examples=[
|
582 |
+
os.path.join(root_path, "person", "men", _)
|
583 |
+
for _ in os.listdir(os.path.join(root_path, "person", "men"))
|
584 |
+
],
|
585 |
+
examples_per_page=4,
|
586 |
+
inputs=image_path_p2p,
|
587 |
+
label="Person Examples ①",
|
588 |
+
)
|
589 |
+
gr.Examples(
|
590 |
+
examples=[
|
591 |
+
os.path.join(root_path, "person", "women", _)
|
592 |
+
for _ in os.listdir(os.path.join(root_path, "person", "women"))
|
593 |
+
],
|
594 |
+
examples_per_page=4,
|
595 |
+
inputs=image_path_p2p,
|
596 |
+
label="Person Examples ②",
|
597 |
+
)
|
598 |
+
gr.Markdown(
|
599 |
+
'<span style="color: #808080; font-size: small;">*Person examples come from the demos of <a href="https://huggingface.co/spaces/levihsu/OOTDiffusion">OOTDiffusion</a> and <a href="https://www.outfitanyone.org">OutfitAnyone</a>. </span>'
|
600 |
+
)
|
601 |
+
with gr.Column():
|
602 |
+
gr.Examples(
|
603 |
+
examples=[
|
604 |
+
os.path.join(root_path, "condition", "upper", _)
|
605 |
+
for _ in os.listdir(os.path.join(root_path, "condition", "upper"))
|
606 |
+
],
|
607 |
+
examples_per_page=4,
|
608 |
+
inputs=cloth_image_p2p,
|
609 |
+
label="Condition Upper Examples",
|
610 |
+
)
|
611 |
+
gr.Examples(
|
612 |
+
examples=[
|
613 |
+
os.path.join(root_path, "condition", "overall", _)
|
614 |
+
for _ in os.listdir(os.path.join(root_path, "condition", "overall"))
|
615 |
+
],
|
616 |
+
examples_per_page=4,
|
617 |
+
inputs=cloth_image_p2p,
|
618 |
+
label="Condition Overall Examples",
|
619 |
+
)
|
620 |
+
condition_person_exm = gr.Examples(
|
621 |
+
examples=[
|
622 |
+
os.path.join(root_path, "condition", "person", _)
|
623 |
+
for _ in os.listdir(os.path.join(root_path, "condition", "person"))
|
624 |
+
],
|
625 |
+
examples_per_page=4,
|
626 |
+
inputs=cloth_image_p2p,
|
627 |
+
label="Condition Reference Person Examples",
|
628 |
+
)
|
629 |
+
gr.Markdown(
|
630 |
+
'<span style="color: #808080; font-size: small;">*Condition examples come from the Internet. </span>'
|
631 |
+
)
|
632 |
+
|
633 |
+
image_path_p2p.change(
|
634 |
+
person_example_fn, inputs=image_path_p2p, outputs=person_image_p2p
|
635 |
+
)
|
636 |
+
|
637 |
+
submit_p2p.click(
|
638 |
+
submit_function_p2p,
|
639 |
+
[
|
640 |
+
person_image_p2p,
|
641 |
+
cloth_image_p2p,
|
642 |
+
num_inference_steps_p2p,
|
643 |
+
guidance_scale_p2p,
|
644 |
+
seed_p2p],
|
645 |
+
result_image_p2p,
|
646 |
+
)
|
647 |
+
|
648 |
+
with gr.Tab("Mask-based & Flux.1 Fill Dev"):
|
649 |
+
with gr.Row():
|
650 |
+
with gr.Column(scale=1, min_width=350):
|
651 |
+
with gr.Row():
|
652 |
+
image_path_flux = gr.Image(
|
653 |
+
type="filepath",
|
654 |
+
interactive=True,
|
655 |
+
visible=False,
|
656 |
)
|
657 |
+
person_image_flux = gr.ImageEditor(
|
658 |
+
interactive=True, label="Person Image", type="filepath"
|
659 |
)
|
660 |
+
|
661 |
+
with gr.Row():
|
662 |
+
with gr.Column(scale=1, min_width=230):
|
663 |
+
cloth_image_flux = gr.Image(
|
664 |
+
interactive=True, label="Condition Image", type="filepath"
|
665 |
+
)
|
666 |
+
with gr.Column(scale=1, min_width=120):
|
667 |
+
gr.Markdown(
|
668 |
+
'<span style="color: #808080; font-size: small;">Two ways to provide Mask:<br>1. Upload the person image and use the `🖌️` above to draw the Mask (higher priority)<br>2. Select the `Try-On Cloth Type` to generate automatically </span>'
|
669 |
+
)
|
670 |
+
cloth_type = gr.Radio(
|
671 |
+
label="Try-On Cloth Type",
|
672 |
+
choices=["upper", "lower", "overall"],
|
673 |
+
value="upper",
|
674 |
+
)
|
675 |
+
|
676 |
+
submit_flux = gr.Button("Submit")
|
677 |
+
gr.Markdown(
|
678 |
+
'<center><span style="color: #FF0000">!!! Click only Once, Wait for Delay !!!</span></center>'
|
679 |
+
)
|
680 |
+
|
681 |
+
with gr.Accordion("Advanced Options", open=False):
|
682 |
+
num_inference_steps_flux = gr.Slider(
|
683 |
+
label="Inference Step", minimum=10, maximum=100, step=5, value=50
|
684 |
)
|
685 |
+
# Guidence Scale
|
686 |
+
guidance_scale_flux = gr.Slider(
|
687 |
+
label="CFG Strenth", minimum=0.0, maximum=7.5, step=0.5, value=2.5
|
|
|
|
|
|
|
|
|
|
|
688 |
)
|
689 |
+
# Random Seed
|
690 |
+
seed_flux = gr.Slider(
|
691 |
+
label="Seed", minimum=-1, maximum=10000, step=1, value=42
|
|
|
|
|
|
|
|
|
|
|
692 |
)
|
693 |
+
show_type = gr.Radio(
|
694 |
+
label="Show Type",
|
695 |
+
choices=["result only", "input & result", "input & mask & result"],
|
696 |
+
value="input & mask & result",
|
697 |
)
|
698 |
|
699 |
+
with gr.Column(scale=2, min_width=500):
|
700 |
+
result_image_flux = gr.Image(interactive=False, label="Result")
|
701 |
+
|
702 |
+
image_path_flux.change(
|
703 |
+
person_example_fn, inputs=image_path_flux, outputs=person_image_flux
|
704 |
+
)
|
705 |
+
|
706 |
+
submit_flux.click(
|
707 |
+
submit_function_flux,
|
708 |
+
[person_image_flux, cloth_image_flux, cloth_type, num_inference_steps_flux, guidance_scale_flux, seed_flux, show_type],
|
709 |
+
result_image_flux,
|
710 |
+
)
|
711 |
+
|
|
|
|
|
|
|
|
|
712 |
demo.queue().launch(share=True, show_error=True)
|
713 |
|
714 |
|
model/flux/pipeline_flux_tryon.py
ADDED
@@ -0,0 +1,499 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
from typing import Any, Callable, Dict, List, Optional, Union
|
3 |
+
|
4 |
+
import numpy as np
|
5 |
+
import torch
|
6 |
+
from diffusers.image_processor import VaeImageProcessor
|
7 |
+
from diffusers.loaders import (
|
8 |
+
FluxLoraLoaderMixin,
|
9 |
+
FromSingleFileMixin,
|
10 |
+
TextualInversionLoaderMixin,
|
11 |
+
)
|
12 |
+
from diffusers.models.autoencoders import AutoencoderKL
|
13 |
+
from diffusers.pipelines.flux.pipeline_flux_fill import (
|
14 |
+
calculate_shift,
|
15 |
+
retrieve_latents,
|
16 |
+
retrieve_timesteps,
|
17 |
+
)
|
18 |
+
from diffusers.pipelines.flux.pipeline_output import FluxPipelineOutput
|
19 |
+
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
|
20 |
+
from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
|
21 |
+
from diffusers.utils import logging
|
22 |
+
from diffusers.utils.torch_utils import randn_tensor
|
23 |
+
|
24 |
+
from model.flux.transformer_flux import FluxTransformer2DModel
|
25 |
+
|
26 |
+
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
27 |
+
|
28 |
+
# Modified from `diffusers.pipelines.flux.pipeline_flux_fill.FluxFillPipeline`
|
29 |
+
class FluxTryOnPipeline(
|
30 |
+
DiffusionPipeline,
|
31 |
+
FluxLoraLoaderMixin,
|
32 |
+
FromSingleFileMixin,
|
33 |
+
TextualInversionLoaderMixin,
|
34 |
+
):
|
35 |
+
model_cpu_offload_seq = "transformer->vae"
|
36 |
+
_optional_components = []
|
37 |
+
_callback_tensor_inputs = ["latents"]
|
38 |
+
|
39 |
+
def __init__(
|
40 |
+
self,
|
41 |
+
vae: AutoencoderKL,
|
42 |
+
scheduler: FlowMatchEulerDiscreteScheduler,
|
43 |
+
transformer: FluxTransformer2DModel,
|
44 |
+
):
|
45 |
+
super().__init__()
|
46 |
+
self.register_modules(
|
47 |
+
vae=vae,
|
48 |
+
scheduler=scheduler,
|
49 |
+
transformer=transformer,
|
50 |
+
)
|
51 |
+
|
52 |
+
self.vae_scale_factor = (
|
53 |
+
2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8
|
54 |
+
)
|
55 |
+
|
56 |
+
# Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible
|
57 |
+
# by the patch size. So the vae scale factor is multiplied by the patch size to account for this
|
58 |
+
self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2)
|
59 |
+
self.mask_processor = VaeImageProcessor(
|
60 |
+
vae_scale_factor=self.vae_scale_factor * 2,
|
61 |
+
vae_latent_channels=self.vae.config.latent_channels,
|
62 |
+
do_normalize=False,
|
63 |
+
do_binarize=True,
|
64 |
+
do_convert_grayscale=True,
|
65 |
+
)
|
66 |
+
self.default_sample_size = 128
|
67 |
+
|
68 |
+
self.transformer.remove_text_layers() # TryOnEdit: remove text layers
|
69 |
+
|
70 |
+
@classmethod
|
71 |
+
def from_pretrained(cls, pretrained_model_name_or_path, subfolder=None, **kwargs):
|
72 |
+
transformer = FluxTransformer2DModel.from_pretrained(pretrained_model_name_or_path, subfolder="transformer")
|
73 |
+
transformer.remove_text_layers()
|
74 |
+
vae = AutoencoderKL.from_pretrained(pretrained_model_name_or_path, subfolder="vae")
|
75 |
+
scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(pretrained_model_name_or_path, subfolder="scheduler")
|
76 |
+
return FluxTryOnPipeline(vae, scheduler, transformer)
|
77 |
+
|
78 |
+
def prepare_mask_latents(
|
79 |
+
self,
|
80 |
+
mask,
|
81 |
+
masked_image,
|
82 |
+
batch_size,
|
83 |
+
num_channels_latents,
|
84 |
+
num_images_per_prompt,
|
85 |
+
height,
|
86 |
+
width,
|
87 |
+
dtype,
|
88 |
+
device,
|
89 |
+
generator,
|
90 |
+
):
|
91 |
+
# 1. calculate the height and width of the latents
|
92 |
+
# VAE applies 8x compression on images but we must also account for packing which requires
|
93 |
+
# latent height and width to be divisible by 2.
|
94 |
+
height = 2 * (int(height) // (self.vae_scale_factor * 2))
|
95 |
+
width = 2 * (int(width) // (self.vae_scale_factor * 2))
|
96 |
+
|
97 |
+
# 2. encode the masked image
|
98 |
+
if masked_image.shape[1] == num_channels_latents:
|
99 |
+
masked_image_latents = masked_image
|
100 |
+
else:
|
101 |
+
masked_image_latents = retrieve_latents(self.vae.encode(masked_image), generator=generator)
|
102 |
+
|
103 |
+
masked_image_latents = (masked_image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor
|
104 |
+
masked_image_latents = masked_image_latents.to(device=device, dtype=dtype)
|
105 |
+
|
106 |
+
# 3. duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method
|
107 |
+
batch_size = batch_size * num_images_per_prompt
|
108 |
+
if mask.shape[0] < batch_size:
|
109 |
+
if not batch_size % mask.shape[0] == 0:
|
110 |
+
raise ValueError(
|
111 |
+
"The passed mask and the required batch size don't match. Masks are supposed to be duplicated to"
|
112 |
+
f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number"
|
113 |
+
" of masks that you pass is divisible by the total requested batch size."
|
114 |
+
)
|
115 |
+
mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1)
|
116 |
+
if masked_image_latents.shape[0] < batch_size:
|
117 |
+
if not batch_size % masked_image_latents.shape[0] == 0:
|
118 |
+
raise ValueError(
|
119 |
+
"The passed images and the required batch size don't match. Images are supposed to be duplicated"
|
120 |
+
f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed."
|
121 |
+
" Make sure the number of images that you pass is divisible by the total requested batch size."
|
122 |
+
)
|
123 |
+
masked_image_latents = masked_image_latents.repeat(batch_size // masked_image_latents.shape[0], 1, 1, 1)
|
124 |
+
|
125 |
+
# 4. pack the masked_image_latents
|
126 |
+
# batch_size, num_channels_latents, height, width -> batch_size, height//2 * width//2 , num_channels_latents*4
|
127 |
+
masked_image_latents = self._pack_latents(
|
128 |
+
masked_image_latents,
|
129 |
+
batch_size,
|
130 |
+
num_channels_latents,
|
131 |
+
height,
|
132 |
+
width,
|
133 |
+
)
|
134 |
+
|
135 |
+
# 5.resize mask to latents shape we we concatenate the mask to the latents
|
136 |
+
mask = mask[:, 0, :, :] # batch_size, 8 * height, 8 * width (mask has not been 8x compressed)
|
137 |
+
mask = mask.view(
|
138 |
+
batch_size, height, self.vae_scale_factor, width, self.vae_scale_factor
|
139 |
+
) # batch_size, height, 8, width, 8
|
140 |
+
mask = mask.permute(0, 2, 4, 1, 3) # batch_size, 8, 8, height, width
|
141 |
+
mask = mask.reshape(
|
142 |
+
batch_size, self.vae_scale_factor * self.vae_scale_factor, height, width
|
143 |
+
) # batch_size, 8*8, height, width
|
144 |
+
|
145 |
+
# 6. pack the mask:
|
146 |
+
# batch_size, 64, height, width -> batch_size, height//2 * width//2 , 64*2*2
|
147 |
+
mask = self._pack_latents(
|
148 |
+
mask,
|
149 |
+
batch_size,
|
150 |
+
self.vae_scale_factor * self.vae_scale_factor,
|
151 |
+
height,
|
152 |
+
width,
|
153 |
+
)
|
154 |
+
mask = mask.to(device=device, dtype=dtype)
|
155 |
+
|
156 |
+
return mask, masked_image_latents
|
157 |
+
|
158 |
+
def check_inputs(
|
159 |
+
self,
|
160 |
+
height,
|
161 |
+
width,
|
162 |
+
callback_on_step_end_tensor_inputs=None,
|
163 |
+
max_sequence_length=None,
|
164 |
+
image=None,
|
165 |
+
mask_image=None,
|
166 |
+
condition_image=None,
|
167 |
+
masked_image_latents=None,
|
168 |
+
):
|
169 |
+
if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0:
|
170 |
+
logger.warning(
|
171 |
+
f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly"
|
172 |
+
)
|
173 |
+
|
174 |
+
if callback_on_step_end_tensor_inputs is not None and not all(
|
175 |
+
k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
|
176 |
+
):
|
177 |
+
raise ValueError(
|
178 |
+
f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
|
179 |
+
)
|
180 |
+
|
181 |
+
if max_sequence_length is not None and max_sequence_length > 512:
|
182 |
+
raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}")
|
183 |
+
|
184 |
+
if image is not None and masked_image_latents is not None:
|
185 |
+
raise ValueError(
|
186 |
+
"Please provide either `image` or `masked_image_latents`, `masked_image_latents` should not be passed."
|
187 |
+
)
|
188 |
+
|
189 |
+
if image is not None and mask_image is None:
|
190 |
+
raise ValueError("Please provide `mask_image` when passing `image`.")
|
191 |
+
|
192 |
+
if condition_image is None:
|
193 |
+
raise ValueError("Please provide `condition_image`.")
|
194 |
+
|
195 |
+
@staticmethod
|
196 |
+
# Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._prepare_latent_image_ids
|
197 |
+
def _prepare_latent_image_ids(batch_size, height, width, device, dtype):
|
198 |
+
latent_image_ids = torch.zeros(height, width, 3)
|
199 |
+
latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height)[:, None]
|
200 |
+
latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width)[None, :]
|
201 |
+
|
202 |
+
latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape
|
203 |
+
|
204 |
+
latent_image_ids = latent_image_ids.reshape(
|
205 |
+
latent_image_id_height * latent_image_id_width, latent_image_id_channels
|
206 |
+
)
|
207 |
+
|
208 |
+
return latent_image_ids.to(device=device, dtype=dtype)
|
209 |
+
|
210 |
+
@staticmethod
|
211 |
+
# Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._pack_latents
|
212 |
+
def _pack_latents(latents, batch_size, num_channels_latents, height, width):
|
213 |
+
latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2)
|
214 |
+
latents = latents.permute(0, 2, 4, 1, 3, 5)
|
215 |
+
latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4)
|
216 |
+
|
217 |
+
return latents
|
218 |
+
|
219 |
+
@staticmethod
|
220 |
+
# Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline._unpack_latents
|
221 |
+
def _unpack_latents(latents, height, width, vae_scale_factor):
|
222 |
+
batch_size, num_patches, channels = latents.shape
|
223 |
+
|
224 |
+
# VAE applies 8x compression on images but we must also account for packing which requires
|
225 |
+
# latent height and width to be divisible by 2.
|
226 |
+
height = 2 * (int(height) // (vae_scale_factor * 2))
|
227 |
+
width = 2 * (int(width) // (vae_scale_factor * 2))
|
228 |
+
|
229 |
+
latents = latents.view(batch_size, height // 2, width // 2, channels // 4, 2, 2)
|
230 |
+
latents = latents.permute(0, 3, 1, 4, 2, 5)
|
231 |
+
|
232 |
+
latents = latents.reshape(batch_size, channels // (2 * 2), height, width)
|
233 |
+
|
234 |
+
return latents
|
235 |
+
|
236 |
+
def enable_vae_slicing(self):
|
237 |
+
r"""
|
238 |
+
Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
|
239 |
+
compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
|
240 |
+
"""
|
241 |
+
self.vae.enable_slicing()
|
242 |
+
|
243 |
+
def disable_vae_slicing(self):
|
244 |
+
r"""
|
245 |
+
Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
|
246 |
+
computing decoding in one step.
|
247 |
+
"""
|
248 |
+
self.vae.disable_slicing()
|
249 |
+
|
250 |
+
def enable_vae_tiling(self):
|
251 |
+
r"""
|
252 |
+
Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
|
253 |
+
compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
|
254 |
+
processing larger images.
|
255 |
+
"""
|
256 |
+
self.vae.enable_tiling()
|
257 |
+
|
258 |
+
def disable_vae_tiling(self):
|
259 |
+
r"""
|
260 |
+
Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
|
261 |
+
computing decoding in one step.
|
262 |
+
"""
|
263 |
+
self.vae.disable_tiling()
|
264 |
+
|
265 |
+
# Copied from diffusers.pipelines.flux.pipeline_flux.FluxPipeline.prepare_latents
|
266 |
+
def prepare_latents(
|
267 |
+
self,
|
268 |
+
batch_size,
|
269 |
+
num_channels_latents,
|
270 |
+
height,
|
271 |
+
width,
|
272 |
+
dtype,
|
273 |
+
device,
|
274 |
+
generator,
|
275 |
+
latents=None,
|
276 |
+
):
|
277 |
+
# VAE applies 8x compression on images but we must also account for packing which requires
|
278 |
+
# latent height and width to be divisible by 2.
|
279 |
+
height = 2 * (int(height) // (self.vae_scale_factor * 2))
|
280 |
+
width = 2 * (int(width) // (self.vae_scale_factor * 2))
|
281 |
+
|
282 |
+
shape = (batch_size, num_channels_latents, height, width)
|
283 |
+
|
284 |
+
if latents is not None:
|
285 |
+
latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype)
|
286 |
+
return latents.to(device=device, dtype=dtype), latent_image_ids
|
287 |
+
|
288 |
+
if isinstance(generator, list) and len(generator) != batch_size:
|
289 |
+
raise ValueError(
|
290 |
+
f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
|
291 |
+
f" size of {batch_size}. Make sure the batch size matches the length of the generators."
|
292 |
+
)
|
293 |
+
|
294 |
+
latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
|
295 |
+
latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width)
|
296 |
+
|
297 |
+
latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype)
|
298 |
+
|
299 |
+
return latents, latent_image_ids
|
300 |
+
|
301 |
+
@property
|
302 |
+
def guidance_scale(self):
|
303 |
+
return self._guidance_scale
|
304 |
+
|
305 |
+
@property
|
306 |
+
def joint_attention_kwargs(self):
|
307 |
+
return self._joint_attention_kwargs
|
308 |
+
|
309 |
+
@property
|
310 |
+
def num_timesteps(self):
|
311 |
+
return self._num_timesteps
|
312 |
+
|
313 |
+
@property
|
314 |
+
def interrupt(self):
|
315 |
+
return self._interrupt
|
316 |
+
|
317 |
+
@torch.no_grad()
|
318 |
+
def __call__(
|
319 |
+
self,
|
320 |
+
image: Optional[torch.FloatTensor] = None,
|
321 |
+
condition_image: Optional[torch.FloatTensor] = None, # TryOnEdit: condition image (garment)
|
322 |
+
mask_image: Optional[torch.FloatTensor] = None,
|
323 |
+
masked_image_latents: Optional[torch.FloatTensor] = None,
|
324 |
+
height: Optional[int] = None,
|
325 |
+
width: Optional[int] = None,
|
326 |
+
num_inference_steps: int = 50,
|
327 |
+
sigmas: Optional[List[float]] = None,
|
328 |
+
guidance_scale: float = 30.0,
|
329 |
+
num_images_per_prompt: Optional[int] = 1,
|
330 |
+
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
331 |
+
latents: Optional[torch.FloatTensor] = None,
|
332 |
+
output_type: Optional[str] = "pil",
|
333 |
+
return_dict: bool = True,
|
334 |
+
joint_attention_kwargs: Optional[Dict[str, Any]] = None,
|
335 |
+
callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
|
336 |
+
callback_on_step_end_tensor_inputs: List[str] = ["latents"],
|
337 |
+
max_sequence_length: int = 512,
|
338 |
+
):
|
339 |
+
height = height or self.default_sample_size * self.vae_scale_factor
|
340 |
+
width = width or self.default_sample_size * self.vae_scale_factor
|
341 |
+
|
342 |
+
# 1. Check inputs. Raise error if not correct
|
343 |
+
self.check_inputs(
|
344 |
+
height,
|
345 |
+
width,
|
346 |
+
callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
|
347 |
+
max_sequence_length=max_sequence_length,
|
348 |
+
image=image,
|
349 |
+
mask_image=mask_image,
|
350 |
+
condition_image=condition_image,
|
351 |
+
masked_image_latents=masked_image_latents,
|
352 |
+
)
|
353 |
+
|
354 |
+
self._guidance_scale = guidance_scale
|
355 |
+
self._joint_attention_kwargs = joint_attention_kwargs
|
356 |
+
self._interrupt = False
|
357 |
+
|
358 |
+
# 2. Define call parameters
|
359 |
+
batch_size = 1
|
360 |
+
device = self._execution_device
|
361 |
+
dtype = self.transformer.dtype
|
362 |
+
|
363 |
+
# 3. Prepare prompt embeddings
|
364 |
+
lora_scale = (
|
365 |
+
self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None
|
366 |
+
)
|
367 |
+
|
368 |
+
# 4. Prepare latent variables
|
369 |
+
num_channels_latents = self.vae.config.latent_channels
|
370 |
+
latents, latent_image_ids = self.prepare_latents(
|
371 |
+
batch_size * num_images_per_prompt,
|
372 |
+
num_channels_latents,
|
373 |
+
height,
|
374 |
+
width * 2, # TryOnEdit: width * 2
|
375 |
+
dtype,
|
376 |
+
device,
|
377 |
+
generator,
|
378 |
+
latents,
|
379 |
+
)
|
380 |
+
|
381 |
+
# 5. Prepare mask and masked image latents
|
382 |
+
if masked_image_latents is not None:
|
383 |
+
masked_image_latents = masked_image_latents.to(latents.device)
|
384 |
+
else:
|
385 |
+
image = self.image_processor.preprocess(image, height=height, width=width)
|
386 |
+
condition_image = self.image_processor.preprocess(condition_image, height=height, width=width)
|
387 |
+
mask_image = self.mask_processor.preprocess(mask_image, height=height, width=width)
|
388 |
+
|
389 |
+
masked_image = image * (1 - mask_image)
|
390 |
+
masked_image = masked_image.to(device=device, dtype=dtype)
|
391 |
+
|
392 |
+
# TryOnEdit: Concat condition image to masked image
|
393 |
+
condition_image = condition_image.to(device=device, dtype=dtype)
|
394 |
+
masked_image = torch.cat((masked_image, condition_image), dim=-1)
|
395 |
+
mask_image = torch.cat((mask_image, torch.zeros_like(mask_image)), dim=-1)
|
396 |
+
|
397 |
+
height, width = image.shape[-2:]
|
398 |
+
mask, masked_image_latents = self.prepare_mask_latents(
|
399 |
+
mask_image,
|
400 |
+
masked_image,
|
401 |
+
batch_size,
|
402 |
+
num_channels_latents,
|
403 |
+
num_images_per_prompt,
|
404 |
+
height,
|
405 |
+
width * 2, # TryOnEdit: width * 2
|
406 |
+
dtype,
|
407 |
+
device,
|
408 |
+
generator,
|
409 |
+
)
|
410 |
+
masked_image_latents = torch.cat((masked_image_latents, mask), dim=-1)
|
411 |
+
|
412 |
+
# 6. Prepare timesteps
|
413 |
+
sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas
|
414 |
+
image_seq_len = latents.shape[1]
|
415 |
+
mu = calculate_shift(
|
416 |
+
image_seq_len,
|
417 |
+
self.scheduler.config.base_image_seq_len,
|
418 |
+
self.scheduler.config.max_image_seq_len,
|
419 |
+
self.scheduler.config.base_shift,
|
420 |
+
self.scheduler.config.max_shift,
|
421 |
+
)
|
422 |
+
timesteps, num_inference_steps = retrieve_timesteps(
|
423 |
+
self.scheduler,
|
424 |
+
num_inference_steps,
|
425 |
+
device,
|
426 |
+
sigmas=sigmas,
|
427 |
+
mu=mu,
|
428 |
+
)
|
429 |
+
num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
|
430 |
+
self._num_timesteps = len(timesteps)
|
431 |
+
|
432 |
+
# handle guidance
|
433 |
+
if self.transformer.config.guidance_embeds:
|
434 |
+
guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32)
|
435 |
+
guidance = guidance.expand(latents.shape[0])
|
436 |
+
else:
|
437 |
+
guidance = None
|
438 |
+
|
439 |
+
# 7. Denoising loop
|
440 |
+
pooled_prompt_embeds = torch.zeros([latents.shape[0], 768], device=device, dtype=dtype) # TryOnEdit: for now, we don't use pooled prompt embeddings
|
441 |
+
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
442 |
+
for i, t in enumerate(timesteps):
|
443 |
+
if self.interrupt:
|
444 |
+
continue
|
445 |
+
|
446 |
+
# broadcast to batch dimension in a way that's compatible with ONNX/Core ML
|
447 |
+
timestep = t.expand(latents.shape[0]).to(latents.dtype)
|
448 |
+
|
449 |
+
noise_pred = self.transformer(
|
450 |
+
hidden_states=torch.cat((latents, masked_image_latents), dim=2),
|
451 |
+
timestep=timestep / 1000,
|
452 |
+
guidance=guidance,
|
453 |
+
pooled_projections=pooled_prompt_embeds,
|
454 |
+
encoder_hidden_states=None,
|
455 |
+
txt_ids=None,
|
456 |
+
img_ids=latent_image_ids,
|
457 |
+
joint_attention_kwargs=self.joint_attention_kwargs,
|
458 |
+
return_dict=False,
|
459 |
+
)[0]
|
460 |
+
|
461 |
+
# compute the previous noisy sample x_t -> x_t-1
|
462 |
+
latents_dtype = latents.dtype
|
463 |
+
latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
|
464 |
+
|
465 |
+
if latents.dtype != latents_dtype:
|
466 |
+
if torch.backends.mps.is_available():
|
467 |
+
# some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
|
468 |
+
latents = latents.to(latents_dtype)
|
469 |
+
|
470 |
+
if callback_on_step_end is not None:
|
471 |
+
callback_kwargs = {}
|
472 |
+
for k in callback_on_step_end_tensor_inputs:
|
473 |
+
callback_kwargs[k] = locals()[k]
|
474 |
+
callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
|
475 |
+
|
476 |
+
latents = callback_outputs.pop("latents", latents)
|
477 |
+
prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
|
478 |
+
|
479 |
+
# call the callback, if provided
|
480 |
+
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
|
481 |
+
progress_bar.update()
|
482 |
+
|
483 |
+
# 8. Post-process the image
|
484 |
+
if output_type == "latent":
|
485 |
+
image = latents
|
486 |
+
else:
|
487 |
+
latents = self._unpack_latents(latents, height, width * 2, self.vae_scale_factor) # TryOnEdit: width * 2
|
488 |
+
latents = latents.split(latents.shape[-1] // 2, dim=-1)[0] # TryOnEdit: split along the last dimension
|
489 |
+
latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor
|
490 |
+
image = self.vae.decode(latents, return_dict=False)[0]
|
491 |
+
image = self.image_processor.postprocess(image, output_type=output_type)
|
492 |
+
|
493 |
+
# Offload all models
|
494 |
+
self.maybe_free_model_hooks()
|
495 |
+
|
496 |
+
if not return_dict:
|
497 |
+
return (image,)
|
498 |
+
|
499 |
+
return FluxPipelineOutput(images=image)
|
model/flux/transformer_flux.py
ADDED
@@ -0,0 +1,672 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Any, Dict, Optional, Tuple, Union
|
2 |
+
|
3 |
+
import numpy as np
|
4 |
+
import torch
|
5 |
+
import torch.nn as nn
|
6 |
+
import torch.nn.functional as F
|
7 |
+
|
8 |
+
from diffusers.models.modeling_utils import ModelMixin
|
9 |
+
from diffusers.models.normalization import AdaLayerNormContinuous, AdaLayerNormZero, AdaLayerNormZeroSingle
|
10 |
+
from diffusers.utils import USE_PEFT_BACKEND, is_torch_version, logging, scale_lora_layers, unscale_lora_layers
|
11 |
+
from diffusers.utils.torch_utils import maybe_allow_in_graph
|
12 |
+
from diffusers.models.embeddings import CombinedTimestepGuidanceTextProjEmbeddings, CombinedTimestepTextProjEmbeddings, FluxPosEmbed
|
13 |
+
from diffusers.models.modeling_outputs import Transformer2DModelOutput
|
14 |
+
|
15 |
+
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
16 |
+
from diffusers.loaders import FromOriginalModelMixin, PeftAdapterMixin
|
17 |
+
from diffusers.models.attention import FeedForward
|
18 |
+
from diffusers.models.attention_processor import (
|
19 |
+
Attention,
|
20 |
+
AttentionProcessor,
|
21 |
+
FusedFluxAttnProcessor2_0,
|
22 |
+
)
|
23 |
+
|
24 |
+
logger = logging.get_logger(__name__) # pylint: disable=invalid-name
|
25 |
+
|
26 |
+
|
27 |
+
# Modified from `diffusers.models.attention_processor.FluxAttnProcessor2_0`
|
28 |
+
class FluxAttnProcessor2_0:
|
29 |
+
"""Attention processor used typically in processing the SD3-like self-attention projections."""
|
30 |
+
|
31 |
+
def __init__(self):
|
32 |
+
if not hasattr(F, "scaled_dot_product_attention"):
|
33 |
+
raise ImportError("FluxAttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
|
34 |
+
|
35 |
+
def __call__(
|
36 |
+
self,
|
37 |
+
attn: Attention,
|
38 |
+
hidden_states: torch.FloatTensor,
|
39 |
+
encoder_hidden_states: torch.FloatTensor = None,
|
40 |
+
attention_mask: Optional[torch.FloatTensor] = None,
|
41 |
+
image_rotary_emb: Optional[torch.Tensor] = None,
|
42 |
+
) -> torch.FloatTensor:
|
43 |
+
batch_size, _, _ = hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
|
44 |
+
|
45 |
+
# `sample` projections.
|
46 |
+
query = attn.to_q(hidden_states)
|
47 |
+
key = attn.to_k(hidden_states)
|
48 |
+
value = attn.to_v(hidden_states)
|
49 |
+
|
50 |
+
inner_dim = key.shape[-1]
|
51 |
+
head_dim = inner_dim // attn.heads
|
52 |
+
|
53 |
+
query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
54 |
+
key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
55 |
+
value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
|
56 |
+
|
57 |
+
if attn.norm_q is not None:
|
58 |
+
query = attn.norm_q(query)
|
59 |
+
if attn.norm_k is not None:
|
60 |
+
key = attn.norm_k(key)
|
61 |
+
|
62 |
+
# the attention in FluxSingleTransformerBlock does not use `encoder_hidden_states`
|
63 |
+
if encoder_hidden_states is not None:
|
64 |
+
# `context` projections.
|
65 |
+
encoder_hidden_states_query_proj = attn.add_q_proj(encoder_hidden_states)
|
66 |
+
encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states)
|
67 |
+
encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states)
|
68 |
+
|
69 |
+
encoder_hidden_states_query_proj = encoder_hidden_states_query_proj.view(
|
70 |
+
batch_size, -1, attn.heads, head_dim
|
71 |
+
).transpose(1, 2)
|
72 |
+
encoder_hidden_states_key_proj = encoder_hidden_states_key_proj.view(
|
73 |
+
batch_size, -1, attn.heads, head_dim
|
74 |
+
).transpose(1, 2)
|
75 |
+
encoder_hidden_states_value_proj = encoder_hidden_states_value_proj.view(
|
76 |
+
batch_size, -1, attn.heads, head_dim
|
77 |
+
).transpose(1, 2)
|
78 |
+
|
79 |
+
if attn.norm_added_q is not None:
|
80 |
+
encoder_hidden_states_query_proj = attn.norm_added_q(encoder_hidden_states_query_proj)
|
81 |
+
if attn.norm_added_k is not None:
|
82 |
+
encoder_hidden_states_key_proj = attn.norm_added_k(encoder_hidden_states_key_proj)
|
83 |
+
|
84 |
+
# attention
|
85 |
+
query = torch.cat([encoder_hidden_states_query_proj, query], dim=2)
|
86 |
+
key = torch.cat([encoder_hidden_states_key_proj, key], dim=2)
|
87 |
+
value = torch.cat([encoder_hidden_states_value_proj, value], dim=2)
|
88 |
+
|
89 |
+
if image_rotary_emb is not None:
|
90 |
+
from diffusers.models.embeddings import apply_rotary_emb
|
91 |
+
|
92 |
+
query = apply_rotary_emb(query, image_rotary_emb)
|
93 |
+
key = apply_rotary_emb(key, image_rotary_emb)
|
94 |
+
|
95 |
+
hidden_states = F.scaled_dot_product_attention(query, key, value, dropout_p=0.0, is_causal=False)
|
96 |
+
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
|
97 |
+
hidden_states = hidden_states.to(query.dtype)
|
98 |
+
|
99 |
+
if encoder_hidden_states is not None:
|
100 |
+
encoder_hidden_states, hidden_states = (
|
101 |
+
hidden_states[:, : encoder_hidden_states.shape[1]],
|
102 |
+
hidden_states[:, encoder_hidden_states.shape[1] :],
|
103 |
+
)
|
104 |
+
encoder_hidden_states = attn.to_add_out(encoder_hidden_states)
|
105 |
+
|
106 |
+
# edited for try-on
|
107 |
+
if not attn.pre_only:
|
108 |
+
# linear proj
|
109 |
+
hidden_states = attn.to_out[0](hidden_states)
|
110 |
+
# dropout
|
111 |
+
hidden_states = attn.to_out[1](hidden_states)
|
112 |
+
|
113 |
+
if encoder_hidden_states is not None:
|
114 |
+
return hidden_states, encoder_hidden_states
|
115 |
+
else:
|
116 |
+
return hidden_states
|
117 |
+
|
118 |
+
|
119 |
+
@maybe_allow_in_graph
|
120 |
+
class FluxSingleTransformerBlock(nn.Module):
|
121 |
+
r"""
|
122 |
+
A Transformer block following the MMDiT architecture, introduced in Stable Diffusion 3.
|
123 |
+
|
124 |
+
Reference: https://arxiv.org/abs/2403.03206
|
125 |
+
|
126 |
+
Parameters:
|
127 |
+
dim (`int`): The number of channels in the input and output.
|
128 |
+
num_attention_heads (`int`): The number of heads to use for multi-head attention.
|
129 |
+
attention_head_dim (`int`): The number of channels in each head.
|
130 |
+
context_pre_only (`bool`): Boolean to determine if we should add some blocks associated with the
|
131 |
+
processing of `context` conditions.
|
132 |
+
"""
|
133 |
+
|
134 |
+
def __init__(self, dim, num_attention_heads, attention_head_dim, mlp_ratio=4.0):
|
135 |
+
super().__init__()
|
136 |
+
self.mlp_hidden_dim = int(dim * mlp_ratio)
|
137 |
+
|
138 |
+
self.norm = AdaLayerNormZeroSingle(dim)
|
139 |
+
self.proj_mlp = nn.Linear(dim, self.mlp_hidden_dim)
|
140 |
+
self.act_mlp = nn.GELU(approximate="tanh")
|
141 |
+
self.proj_out = nn.Linear(dim + self.mlp_hidden_dim, dim)
|
142 |
+
|
143 |
+
processor = FluxAttnProcessor2_0()
|
144 |
+
self.attn = Attention(
|
145 |
+
query_dim=dim,
|
146 |
+
cross_attention_dim=None,
|
147 |
+
dim_head=attention_head_dim,
|
148 |
+
heads=num_attention_heads,
|
149 |
+
out_dim=dim,
|
150 |
+
bias=True,
|
151 |
+
processor=processor,
|
152 |
+
qk_norm="rms_norm",
|
153 |
+
eps=1e-6,
|
154 |
+
pre_only=True,
|
155 |
+
)
|
156 |
+
|
157 |
+
def forward(
|
158 |
+
self,
|
159 |
+
hidden_states: torch.FloatTensor,
|
160 |
+
temb: torch.FloatTensor,
|
161 |
+
image_rotary_emb=None,
|
162 |
+
joint_attention_kwargs=None,
|
163 |
+
):
|
164 |
+
residual = hidden_states
|
165 |
+
norm_hidden_states, gate = self.norm(hidden_states, emb=temb)
|
166 |
+
mlp_hidden_states = self.act_mlp(self.proj_mlp(norm_hidden_states))
|
167 |
+
joint_attention_kwargs = joint_attention_kwargs or {}
|
168 |
+
attn_output = self.attn(
|
169 |
+
hidden_states=norm_hidden_states,
|
170 |
+
image_rotary_emb=image_rotary_emb,
|
171 |
+
**joint_attention_kwargs,
|
172 |
+
)
|
173 |
+
|
174 |
+
hidden_states = torch.cat([attn_output, mlp_hidden_states], dim=2)
|
175 |
+
gate = gate.unsqueeze(1)
|
176 |
+
hidden_states = gate * self.proj_out(hidden_states)
|
177 |
+
hidden_states = residual + hidden_states
|
178 |
+
if hidden_states.dtype == torch.float16:
|
179 |
+
hidden_states = hidden_states.clip(-65504, 65504)
|
180 |
+
|
181 |
+
return hidden_states
|
182 |
+
|
183 |
+
|
184 |
+
@maybe_allow_in_graph
|
185 |
+
class FluxTransformerBlock(nn.Module):
|
186 |
+
r"""
|
187 |
+
A Transformer block following the MMDiT architecture, introduced in Stable Diffusion 3.
|
188 |
+
|
189 |
+
Reference: https://arxiv.org/abs/2403.03206
|
190 |
+
|
191 |
+
Parameters:
|
192 |
+
dim (`int`): The number of channels in the input and output.
|
193 |
+
num_attention_heads (`int`): The number of heads to use for multi-head attention.
|
194 |
+
attention_head_dim (`int`): The number of channels in each head.
|
195 |
+
context_pre_only (`bool`): Boolean to determine if we should add some blocks associated with the
|
196 |
+
processing of `context` conditions.
|
197 |
+
"""
|
198 |
+
|
199 |
+
def __init__(self, dim, num_attention_heads, attention_head_dim, qk_norm="rms_norm", eps=1e-6):
|
200 |
+
super().__init__()
|
201 |
+
|
202 |
+
self.norm1 = AdaLayerNormZero(dim)
|
203 |
+
|
204 |
+
self.norm1_context = AdaLayerNormZero(dim)
|
205 |
+
|
206 |
+
if hasattr(F, "scaled_dot_product_attention"):
|
207 |
+
processor = FluxAttnProcessor2_0()
|
208 |
+
else:
|
209 |
+
raise ValueError(
|
210 |
+
"The current PyTorch version does not support the `scaled_dot_product_attention` function."
|
211 |
+
)
|
212 |
+
self.attn = Attention(
|
213 |
+
query_dim=dim,
|
214 |
+
cross_attention_dim=None,
|
215 |
+
added_kv_proj_dim=dim,
|
216 |
+
dim_head=attention_head_dim,
|
217 |
+
heads=num_attention_heads,
|
218 |
+
out_dim=dim,
|
219 |
+
context_pre_only=False,
|
220 |
+
bias=True,
|
221 |
+
processor=processor,
|
222 |
+
qk_norm=qk_norm,
|
223 |
+
eps=eps,
|
224 |
+
)
|
225 |
+
|
226 |
+
self.norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
|
227 |
+
self.ff = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")
|
228 |
+
|
229 |
+
self.norm2_context = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
|
230 |
+
self.ff_context = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")
|
231 |
+
|
232 |
+
# let chunk size default to None
|
233 |
+
self._chunk_size = None
|
234 |
+
self._chunk_dim = 0
|
235 |
+
|
236 |
+
def remove_text_layers(self):
|
237 |
+
# for try-on, we don't need the text conditioning
|
238 |
+
self.norm1_context = None
|
239 |
+
self.ff_context = None
|
240 |
+
self.norm2_context = None
|
241 |
+
self.attn.to_added_qkv = None
|
242 |
+
self.attn.norm_added_q = None
|
243 |
+
self.attn.norm_added_k = None
|
244 |
+
|
245 |
+
def forward(
|
246 |
+
self,
|
247 |
+
hidden_states: torch.FloatTensor,
|
248 |
+
encoder_hidden_states: torch.FloatTensor,
|
249 |
+
temb: torch.FloatTensor,
|
250 |
+
image_rotary_emb=None,
|
251 |
+
joint_attention_kwargs=None,
|
252 |
+
):
|
253 |
+
norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(hidden_states, emb=temb)
|
254 |
+
|
255 |
+
if encoder_hidden_states is not None:
|
256 |
+
norm_encoder_hidden_states, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp = self.norm1_context(
|
257 |
+
encoder_hidden_states, emb=temb
|
258 |
+
)
|
259 |
+
else:
|
260 |
+
norm_encoder_hidden_states = None
|
261 |
+
|
262 |
+
joint_attention_kwargs = joint_attention_kwargs or {}
|
263 |
+
# Attention.
|
264 |
+
|
265 |
+
outputs = self.attn(
|
266 |
+
hidden_states=norm_hidden_states,
|
267 |
+
encoder_hidden_states=norm_encoder_hidden_states,
|
268 |
+
image_rotary_emb=image_rotary_emb,
|
269 |
+
**joint_attention_kwargs,
|
270 |
+
)
|
271 |
+
if isinstance(outputs, tuple):
|
272 |
+
attn_output, context_attn_output = outputs
|
273 |
+
else:
|
274 |
+
attn_output = outputs
|
275 |
+
|
276 |
+
# Process attention outputs for the `hidden_states`.
|
277 |
+
attn_output = gate_msa.unsqueeze(1) * attn_output
|
278 |
+
hidden_states = hidden_states + attn_output
|
279 |
+
|
280 |
+
norm_hidden_states = self.norm2(hidden_states)
|
281 |
+
norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
|
282 |
+
|
283 |
+
ff_output = self.ff(norm_hidden_states)
|
284 |
+
ff_output = gate_mlp.unsqueeze(1) * ff_output
|
285 |
+
|
286 |
+
hidden_states = hidden_states + ff_output
|
287 |
+
|
288 |
+
# Process attention outputs for the `encoder_hidden_states`.
|
289 |
+
if encoder_hidden_states is not None:
|
290 |
+
context_attn_output = c_gate_msa.unsqueeze(1) * context_attn_output
|
291 |
+
encoder_hidden_states = encoder_hidden_states + context_attn_output
|
292 |
+
|
293 |
+
norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states)
|
294 |
+
norm_encoder_hidden_states = norm_encoder_hidden_states * (1 + c_scale_mlp[:, None]) + c_shift_mlp[:, None]
|
295 |
+
|
296 |
+
context_ff_output = self.ff_context(norm_encoder_hidden_states)
|
297 |
+
encoder_hidden_states = encoder_hidden_states + c_gate_mlp.unsqueeze(1) * context_ff_output
|
298 |
+
if encoder_hidden_states.dtype == torch.float16:
|
299 |
+
encoder_hidden_states = encoder_hidden_states.clip(-65504, 65504)
|
300 |
+
|
301 |
+
return encoder_hidden_states, hidden_states
|
302 |
+
|
303 |
+
|
304 |
+
class FluxTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin, FromOriginalModelMixin):
|
305 |
+
"""
|
306 |
+
The Transformer model introduced in Flux.
|
307 |
+
|
308 |
+
Reference: https://blackforestlabs.ai/announcing-black-forest-labs/
|
309 |
+
|
310 |
+
Parameters:
|
311 |
+
patch_size (`int`): Patch size to turn the input data into small patches.
|
312 |
+
in_channels (`int`, *optional*, defaults to 16): The number of channels in the input.
|
313 |
+
num_layers (`int`, *optional*, defaults to 18): The number of layers of MMDiT blocks to use.
|
314 |
+
num_single_layers (`int`, *optional*, defaults to 18): The number of layers of single DiT blocks to use.
|
315 |
+
attention_head_dim (`int`, *optional*, defaults to 64): The number of channels in each head.
|
316 |
+
num_attention_heads (`int`, *optional*, defaults to 18): The number of heads to use for multi-head attention.
|
317 |
+
joint_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
|
318 |
+
pooled_projection_dim (`int`): Number of dimensions to use when projecting the `pooled_projections`.
|
319 |
+
guidance_embeds (`bool`, defaults to False): Whether to use guidance embeddings.
|
320 |
+
"""
|
321 |
+
|
322 |
+
_supports_gradient_checkpointing = True
|
323 |
+
_no_split_modules = ["FluxTransformerBlock", "FluxSingleTransformerBlock"]
|
324 |
+
|
325 |
+
@register_to_config
|
326 |
+
def __init__(
|
327 |
+
self,
|
328 |
+
patch_size: int = 1,
|
329 |
+
in_channels: int = 64,
|
330 |
+
out_channels: Optional[int] = None,
|
331 |
+
num_layers: int = 19,
|
332 |
+
num_single_layers: int = 38,
|
333 |
+
attention_head_dim: int = 128,
|
334 |
+
num_attention_heads: int = 24,
|
335 |
+
joint_attention_dim: int = 4096,
|
336 |
+
pooled_projection_dim: int = 768,
|
337 |
+
guidance_embeds: bool = False,
|
338 |
+
axes_dims_rope: Tuple[int] = (16, 56, 56),
|
339 |
+
):
|
340 |
+
super().__init__()
|
341 |
+
self.out_channels = out_channels or in_channels
|
342 |
+
self.inner_dim = self.config.num_attention_heads * self.config.attention_head_dim
|
343 |
+
|
344 |
+
self.pos_embed = FluxPosEmbed(theta=10000, axes_dim=axes_dims_rope)
|
345 |
+
|
346 |
+
text_time_guidance_cls = (
|
347 |
+
CombinedTimestepGuidanceTextProjEmbeddings if guidance_embeds else CombinedTimestepTextProjEmbeddings
|
348 |
+
)
|
349 |
+
self.time_text_embed = text_time_guidance_cls(
|
350 |
+
embedding_dim=self.inner_dim, pooled_projection_dim=self.config.pooled_projection_dim
|
351 |
+
)
|
352 |
+
|
353 |
+
self.context_embedder = nn.Linear(self.config.joint_attention_dim, self.inner_dim)
|
354 |
+
self.x_embedder = torch.nn.Linear(self.config.in_channels, self.inner_dim)
|
355 |
+
|
356 |
+
self.transformer_blocks = nn.ModuleList(
|
357 |
+
[
|
358 |
+
FluxTransformerBlock(
|
359 |
+
dim=self.inner_dim,
|
360 |
+
num_attention_heads=self.config.num_attention_heads,
|
361 |
+
attention_head_dim=self.config.attention_head_dim,
|
362 |
+
)
|
363 |
+
for i in range(self.config.num_layers)
|
364 |
+
]
|
365 |
+
)
|
366 |
+
|
367 |
+
self.single_transformer_blocks = nn.ModuleList(
|
368 |
+
[
|
369 |
+
FluxSingleTransformerBlock(
|
370 |
+
dim=self.inner_dim,
|
371 |
+
num_attention_heads=self.config.num_attention_heads,
|
372 |
+
attention_head_dim=self.config.attention_head_dim,
|
373 |
+
)
|
374 |
+
for i in range(self.config.num_single_layers)
|
375 |
+
]
|
376 |
+
)
|
377 |
+
|
378 |
+
self.norm_out = AdaLayerNormContinuous(self.inner_dim, self.inner_dim, elementwise_affine=False, eps=1e-6)
|
379 |
+
self.proj_out = nn.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True)
|
380 |
+
|
381 |
+
self.gradient_checkpointing = False
|
382 |
+
|
383 |
+
@property
|
384 |
+
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors
|
385 |
+
def attn_processors(self) -> Dict[str, AttentionProcessor]:
|
386 |
+
r"""
|
387 |
+
Returns:
|
388 |
+
`dict` of attention processors: A dictionary containing all attention processors used in the model with
|
389 |
+
indexed by its weight name.
|
390 |
+
"""
|
391 |
+
# set recursively
|
392 |
+
processors = {}
|
393 |
+
|
394 |
+
def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
|
395 |
+
if hasattr(module, "get_processor"):
|
396 |
+
processors[f"{name}.processor"] = module.get_processor()
|
397 |
+
|
398 |
+
for sub_name, child in module.named_children():
|
399 |
+
fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
|
400 |
+
|
401 |
+
return processors
|
402 |
+
|
403 |
+
for name, module in self.named_children():
|
404 |
+
fn_recursive_add_processors(name, module, processors)
|
405 |
+
|
406 |
+
return processors
|
407 |
+
|
408 |
+
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor
|
409 |
+
def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
|
410 |
+
r"""
|
411 |
+
Sets the attention processor to use to compute attention.
|
412 |
+
|
413 |
+
Parameters:
|
414 |
+
processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
|
415 |
+
The instantiated processor class or a dictionary of processor classes that will be set as the processor
|
416 |
+
for **all** `Attention` layers.
|
417 |
+
|
418 |
+
If `processor` is a dict, the key needs to define the path to the corresponding cross attention
|
419 |
+
processor. This is strongly recommended when setting trainable attention processors.
|
420 |
+
|
421 |
+
"""
|
422 |
+
count = len(self.attn_processors.keys())
|
423 |
+
|
424 |
+
if isinstance(processor, dict) and len(processor) != count:
|
425 |
+
raise ValueError(
|
426 |
+
f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
|
427 |
+
f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
|
428 |
+
)
|
429 |
+
|
430 |
+
def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
|
431 |
+
if hasattr(module, "set_processor"):
|
432 |
+
if not isinstance(processor, dict):
|
433 |
+
module.set_processor(processor)
|
434 |
+
else:
|
435 |
+
module.set_processor(processor.pop(f"{name}.processor"))
|
436 |
+
|
437 |
+
for sub_name, child in module.named_children():
|
438 |
+
fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
|
439 |
+
|
440 |
+
for name, module in self.named_children():
|
441 |
+
fn_recursive_attn_processor(name, module, processor)
|
442 |
+
|
443 |
+
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections with FusedAttnProcessor2_0->FusedFluxAttnProcessor2_0
|
444 |
+
def fuse_qkv_projections(self):
|
445 |
+
"""
|
446 |
+
Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
|
447 |
+
are fused. For cross-attention modules, key and value projection matrices are fused.
|
448 |
+
|
449 |
+
<Tip warning={true}>
|
450 |
+
|
451 |
+
This API is 🧪 experimental.
|
452 |
+
|
453 |
+
</Tip>
|
454 |
+
"""
|
455 |
+
self.original_attn_processors = None
|
456 |
+
|
457 |
+
for _, attn_processor in self.attn_processors.items():
|
458 |
+
if "Added" in str(attn_processor.__class__.__name__):
|
459 |
+
raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
|
460 |
+
|
461 |
+
self.original_attn_processors = self.attn_processors
|
462 |
+
|
463 |
+
for module in self.modules():
|
464 |
+
if isinstance(module, Attention):
|
465 |
+
module.fuse_projections(fuse=True)
|
466 |
+
|
467 |
+
self.set_attn_processor(FusedFluxAttnProcessor2_0())
|
468 |
+
|
469 |
+
# Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections
|
470 |
+
def unfuse_qkv_projections(self):
|
471 |
+
"""Disables the fused QKV projection if enabled.
|
472 |
+
|
473 |
+
<Tip warning={true}>
|
474 |
+
|
475 |
+
This API is 🧪 experimental.
|
476 |
+
|
477 |
+
</Tip>
|
478 |
+
|
479 |
+
"""
|
480 |
+
if self.original_attn_processors is not None:
|
481 |
+
self.set_attn_processor(self.original_attn_processors)
|
482 |
+
|
483 |
+
def _set_gradient_checkpointing(self, module, value=False):
|
484 |
+
if hasattr(module, "gradient_checkpointing"):
|
485 |
+
module.gradient_checkpointing = value
|
486 |
+
|
487 |
+
def remove_text_layers(self):
|
488 |
+
self.context_embedder = None
|
489 |
+
for transformer_block in self.transformer_blocks:
|
490 |
+
transformer_block.remove_text_layers()
|
491 |
+
|
492 |
+
def forward(
|
493 |
+
self,
|
494 |
+
hidden_states: torch.Tensor,
|
495 |
+
encoder_hidden_states: torch.Tensor = None,
|
496 |
+
condition_hidden_states: torch.Tensor = None,
|
497 |
+
pooled_projections: torch.Tensor = None,
|
498 |
+
timestep: torch.LongTensor = None,
|
499 |
+
img_ids: torch.Tensor = None,
|
500 |
+
txt_ids: torch.Tensor = None,
|
501 |
+
guidance: torch.Tensor = None,
|
502 |
+
joint_attention_kwargs: Optional[Dict[str, Any]] = None,
|
503 |
+
controlnet_block_samples=None,
|
504 |
+
controlnet_single_block_samples=None,
|
505 |
+
return_dict: bool = True,
|
506 |
+
controlnet_blocks_repeat: bool = False,
|
507 |
+
) -> Union[torch.FloatTensor, Transformer2DModelOutput]:
|
508 |
+
"""
|
509 |
+
The [`FluxTransformer2DModel`] forward method.
|
510 |
+
|
511 |
+
Args:
|
512 |
+
hidden_states (`torch.FloatTensor` of shape `(batch size, channel, height, width)`):
|
513 |
+
Input `hidden_states`.
|
514 |
+
encoder_hidden_states (`torch.FloatTensor` of shape `(batch size, sequence_len, embed_dims)`):
|
515 |
+
Conditional embeddings (embeddings computed from the input conditions such as prompts) to use.
|
516 |
+
pooled_projections (`torch.FloatTensor` of shape `(batch_size, projection_dim)`): Embeddings projected
|
517 |
+
from the embeddings of input conditions.
|
518 |
+
timestep ( `torch.LongTensor`):
|
519 |
+
Used to indicate denoising step.
|
520 |
+
block_controlnet_hidden_states: (`list` of `torch.Tensor`):
|
521 |
+
A list of tensors that if specified are added to the residuals of transformer blocks.
|
522 |
+
joint_attention_kwargs (`dict`, *optional*):
|
523 |
+
A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
|
524 |
+
`self.processor` in
|
525 |
+
[diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
|
526 |
+
return_dict (`bool`, *optional*, defaults to `True`):
|
527 |
+
Whether or not to return a [`~models.transformer_2d.Transformer2DModelOutput`] instead of a plain
|
528 |
+
tuple.
|
529 |
+
|
530 |
+
Returns:
|
531 |
+
If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
|
532 |
+
`tuple` where the first element is the sample tensor.
|
533 |
+
"""
|
534 |
+
if joint_attention_kwargs is not None:
|
535 |
+
joint_attention_kwargs = joint_attention_kwargs.copy()
|
536 |
+
lora_scale = joint_attention_kwargs.pop("scale", 1.0)
|
537 |
+
else:
|
538 |
+
lora_scale = 1.0
|
539 |
+
|
540 |
+
if USE_PEFT_BACKEND:
|
541 |
+
# weight the lora layers by setting `lora_scale` for each PEFT layer
|
542 |
+
scale_lora_layers(self, lora_scale)
|
543 |
+
else:
|
544 |
+
if joint_attention_kwargs is not None and joint_attention_kwargs.get("scale", None) is not None:
|
545 |
+
logger.warning(
|
546 |
+
"Passing `scale` via `joint_attention_kwargs` when not using the PEFT backend is ineffective."
|
547 |
+
)
|
548 |
+
|
549 |
+
hidden_states = self.x_embedder(hidden_states)
|
550 |
+
|
551 |
+
timestep = timestep.to(hidden_states.dtype) * 1000
|
552 |
+
guidance = guidance.to(hidden_states.dtype) * 1000 if guidance is not None else None
|
553 |
+
|
554 |
+
temb = self.time_text_embed(timestep, pooled_projections) if guidance is None else self.time_text_embed(timestep, guidance, pooled_projections)
|
555 |
+
|
556 |
+
if encoder_hidden_states is not None:
|
557 |
+
encoder_hidden_states = self.context_embedder(encoder_hidden_states)
|
558 |
+
|
559 |
+
ids = torch.cat((txt_ids, img_ids), dim=0) if txt_ids is not None else img_ids # for try-on, we don't need txt_ids
|
560 |
+
image_rotary_emb = self.pos_embed(ids)
|
561 |
+
|
562 |
+
# MMDiT Blocks
|
563 |
+
for index_block, block in enumerate(self.transformer_blocks):
|
564 |
+
if self.training and self.gradient_checkpointing:
|
565 |
+
def create_custom_forward(module, return_dict=None):
|
566 |
+
def custom_forward(*inputs):
|
567 |
+
if return_dict is not None:
|
568 |
+
return module(*inputs, return_dict=return_dict)
|
569 |
+
else:
|
570 |
+
return module(*inputs)
|
571 |
+
|
572 |
+
return custom_forward
|
573 |
+
|
574 |
+
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
|
575 |
+
result = torch.utils.checkpoint.checkpoint(
|
576 |
+
create_custom_forward(block),
|
577 |
+
hidden_states,
|
578 |
+
encoder_hidden_states,
|
579 |
+
temb,
|
580 |
+
image_rotary_emb,
|
581 |
+
**ckpt_kwargs,
|
582 |
+
)
|
583 |
+
if isinstance(result, tuple):
|
584 |
+
encoder_hidden_states, hidden_states = result
|
585 |
+
else:
|
586 |
+
hidden_states = result
|
587 |
+
|
588 |
+
else:
|
589 |
+
result = block(
|
590 |
+
hidden_states=hidden_states,
|
591 |
+
encoder_hidden_states=encoder_hidden_states,
|
592 |
+
temb=temb,
|
593 |
+
image_rotary_emb=image_rotary_emb,
|
594 |
+
joint_attention_kwargs=joint_attention_kwargs,
|
595 |
+
)
|
596 |
+
if isinstance(result, tuple):
|
597 |
+
encoder_hidden_states, hidden_states = result
|
598 |
+
else:
|
599 |
+
hidden_states = result
|
600 |
+
|
601 |
+
# Condition residual (for try-on pose conditioning)
|
602 |
+
if condition_hidden_states is not None and index_block == 0:
|
603 |
+
hidden_states = hidden_states + condition_hidden_states
|
604 |
+
|
605 |
+
# controlnet residual
|
606 |
+
if controlnet_block_samples is not None:
|
607 |
+
interval_control = len(self.transformer_blocks) / len(controlnet_block_samples)
|
608 |
+
interval_control = int(np.ceil(interval_control))
|
609 |
+
# For Xlabs ControlNet.
|
610 |
+
if controlnet_blocks_repeat:
|
611 |
+
hidden_states = (
|
612 |
+
hidden_states + controlnet_block_samples[index_block % len(controlnet_block_samples)]
|
613 |
+
)
|
614 |
+
else:
|
615 |
+
hidden_states = hidden_states + controlnet_block_samples[index_block // interval_control]
|
616 |
+
|
617 |
+
if encoder_hidden_states is not None:
|
618 |
+
hidden_states = torch.cat([encoder_hidden_states, hidden_states], dim=1)
|
619 |
+
|
620 |
+
# Single DiT Blocks
|
621 |
+
for index_block, block in enumerate(self.single_transformer_blocks):
|
622 |
+
if self.training and self.gradient_checkpointing:
|
623 |
+
|
624 |
+
def create_custom_forward(module, return_dict=None):
|
625 |
+
def custom_forward(*inputs):
|
626 |
+
if return_dict is not None:
|
627 |
+
return module(*inputs, return_dict=return_dict)
|
628 |
+
else:
|
629 |
+
return module(*inputs)
|
630 |
+
|
631 |
+
return custom_forward
|
632 |
+
|
633 |
+
ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
|
634 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
635 |
+
create_custom_forward(block),
|
636 |
+
hidden_states,
|
637 |
+
temb,
|
638 |
+
image_rotary_emb,
|
639 |
+
**ckpt_kwargs,
|
640 |
+
)
|
641 |
+
|
642 |
+
else:
|
643 |
+
hidden_states = block(
|
644 |
+
hidden_states=hidden_states,
|
645 |
+
temb=temb,
|
646 |
+
image_rotary_emb=image_rotary_emb,
|
647 |
+
joint_attention_kwargs=joint_attention_kwargs,
|
648 |
+
)
|
649 |
+
|
650 |
+
# controlnet residual
|
651 |
+
if controlnet_single_block_samples is not None:
|
652 |
+
interval_control = len(self.single_transformer_blocks) / len(controlnet_single_block_samples)
|
653 |
+
interval_control = int(np.ceil(interval_control))
|
654 |
+
hidden_states[:, encoder_hidden_states.shape[1] :, ...] = (
|
655 |
+
hidden_states[:, encoder_hidden_states.shape[1] :, ...]
|
656 |
+
+ controlnet_single_block_samples[index_block // interval_control]
|
657 |
+
)
|
658 |
+
|
659 |
+
if encoder_hidden_states is not None:
|
660 |
+
hidden_states = hidden_states[:, encoder_hidden_states.shape[1] :, ...]
|
661 |
+
|
662 |
+
hidden_states = self.norm_out(hidden_states, temb)
|
663 |
+
output = self.proj_out(hidden_states)
|
664 |
+
|
665 |
+
if USE_PEFT_BACKEND:
|
666 |
+
# remove `lora_scale` from each PEFT layer
|
667 |
+
unscale_lora_layers(self, lora_scale)
|
668 |
+
|
669 |
+
if not return_dict:
|
670 |
+
return (output,)
|
671 |
+
|
672 |
+
return Transformer2DModelOutput(sample=output)
|
model/pipeline.py
CHANGED
@@ -213,3 +213,120 @@ class CatVTONPipeline:
|
|
213 |
if not_safe:
|
214 |
image[i] = nsfw_image
|
215 |
return image
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
213 |
if not_safe:
|
214 |
image[i] = nsfw_image
|
215 |
return image
|
216 |
+
|
217 |
+
|
218 |
+
class CatVTONPix2PixPipeline(CatVTONPipeline):
|
219 |
+
def auto_attn_ckpt_load(self, attn_ckpt, version):
|
220 |
+
# TODO: Temperal fix for the model version
|
221 |
+
if os.path.exists(attn_ckpt):
|
222 |
+
load_checkpoint_in_model(self.attn_modules, os.path.join(attn_ckpt, version, 'attention'))
|
223 |
+
else:
|
224 |
+
repo_path = snapshot_download(repo_id=attn_ckpt)
|
225 |
+
print(f"Downloaded {attn_ckpt} to {repo_path}")
|
226 |
+
load_checkpoint_in_model(self.attn_modules, os.path.join(repo_path, version, 'attention'))
|
227 |
+
|
228 |
+
def check_inputs(self, image, condition_image, width, height):
|
229 |
+
if isinstance(image, torch.Tensor) and isinstance(condition_image, torch.Tensor) and isinstance(torch.Tensor):
|
230 |
+
return image, condition_image
|
231 |
+
image = resize_and_crop(image, (width, height))
|
232 |
+
condition_image = resize_and_padding(condition_image, (width, height))
|
233 |
+
return image, condition_image
|
234 |
+
|
235 |
+
@torch.no_grad()
|
236 |
+
def __call__(
|
237 |
+
self,
|
238 |
+
image: Union[PIL.Image.Image, torch.Tensor],
|
239 |
+
condition_image: Union[PIL.Image.Image, torch.Tensor],
|
240 |
+
num_inference_steps: int = 50,
|
241 |
+
guidance_scale: float = 2.5,
|
242 |
+
height: int = 1024,
|
243 |
+
width: int = 768,
|
244 |
+
generator=None,
|
245 |
+
eta=1.0,
|
246 |
+
**kwargs
|
247 |
+
):
|
248 |
+
concat_dim = -1
|
249 |
+
# Prepare inputs to Tensor
|
250 |
+
image, condition_image = self.check_inputs(image, condition_image, width, height)
|
251 |
+
image = prepare_image(image).to(self.device, dtype=self.weight_dtype)
|
252 |
+
condition_image = prepare_image(condition_image).to(self.device, dtype=self.weight_dtype)
|
253 |
+
# VAE encoding
|
254 |
+
image_latent = compute_vae_encodings(image, self.vae)
|
255 |
+
condition_latent = compute_vae_encodings(condition_image, self.vae)
|
256 |
+
del image, condition_image
|
257 |
+
# Concatenate latents
|
258 |
+
condition_latent_concat = torch.cat([image_latent, condition_latent], dim=concat_dim)
|
259 |
+
# Prepare noise
|
260 |
+
latents = randn_tensor(
|
261 |
+
condition_latent_concat.shape,
|
262 |
+
generator=generator,
|
263 |
+
device=condition_latent_concat.device,
|
264 |
+
dtype=self.weight_dtype,
|
265 |
+
)
|
266 |
+
# Prepare timesteps
|
267 |
+
self.noise_scheduler.set_timesteps(num_inference_steps, device=self.device)
|
268 |
+
timesteps = self.noise_scheduler.timesteps
|
269 |
+
latents = latents * self.noise_scheduler.init_noise_sigma
|
270 |
+
# Classifier-Free Guidance
|
271 |
+
if do_classifier_free_guidance := (guidance_scale > 1.0):
|
272 |
+
condition_latent_concat = torch.cat(
|
273 |
+
[
|
274 |
+
torch.cat([image_latent, torch.zeros_like(condition_latent)], dim=concat_dim),
|
275 |
+
condition_latent_concat,
|
276 |
+
]
|
277 |
+
)
|
278 |
+
|
279 |
+
# Denoising loop
|
280 |
+
extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
|
281 |
+
num_warmup_steps = (len(timesteps) - num_inference_steps * self.noise_scheduler.order)
|
282 |
+
with tqdm.tqdm(total=num_inference_steps) as progress_bar:
|
283 |
+
for i, t in enumerate(timesteps):
|
284 |
+
# expand the latents if we are doing classifier free guidance
|
285 |
+
latent_model_input = (torch.cat([latents] * 2) if do_classifier_free_guidance else latents)
|
286 |
+
latent_model_input = self.noise_scheduler.scale_model_input(latent_model_input, t)
|
287 |
+
# prepare the input for the inpainting model
|
288 |
+
p2p_latent_model_input = torch.cat([latent_model_input, condition_latent_concat], dim=1)
|
289 |
+
# predict the noise residual
|
290 |
+
noise_pred= self.unet(
|
291 |
+
p2p_latent_model_input,
|
292 |
+
t.to(self.device),
|
293 |
+
encoder_hidden_states=None,
|
294 |
+
return_dict=False,
|
295 |
+
)[0]
|
296 |
+
# perform guidance
|
297 |
+
if do_classifier_free_guidance:
|
298 |
+
noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
|
299 |
+
noise_pred = noise_pred_uncond + guidance_scale * (
|
300 |
+
noise_pred_text - noise_pred_uncond
|
301 |
+
)
|
302 |
+
# compute the previous noisy sample x_t -> x_t-1
|
303 |
+
latents = self.noise_scheduler.step(
|
304 |
+
noise_pred, t, latents, **extra_step_kwargs
|
305 |
+
).prev_sample
|
306 |
+
# call the callback, if provided
|
307 |
+
if i == len(timesteps) - 1 or (
|
308 |
+
(i + 1) > num_warmup_steps
|
309 |
+
and (i + 1) % self.noise_scheduler.order == 0
|
310 |
+
):
|
311 |
+
progress_bar.update()
|
312 |
+
|
313 |
+
# Decode the final latents
|
314 |
+
latents = latents.split(latents.shape[concat_dim] // 2, dim=concat_dim)[0]
|
315 |
+
latents = 1 / self.vae.config.scaling_factor * latents
|
316 |
+
image = self.vae.decode(latents.to(self.device, dtype=self.weight_dtype)).sample
|
317 |
+
image = (image / 2 + 0.5).clamp(0, 1)
|
318 |
+
# we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
|
319 |
+
image = image.cpu().permute(0, 2, 3, 1).float().numpy()
|
320 |
+
image = numpy_to_pil(image)
|
321 |
+
|
322 |
+
# Safety Check
|
323 |
+
if not self.skip_safety_check:
|
324 |
+
current_script_directory = os.path.dirname(os.path.realpath(__file__))
|
325 |
+
nsfw_image = os.path.join(os.path.dirname(current_script_directory), 'resource', 'img', 'NSFW.jpg')
|
326 |
+
nsfw_image = PIL.Image.open(nsfw_image).resize(image[0].size)
|
327 |
+
image_np = np.array(image)
|
328 |
+
_, has_nsfw_concept = self.run_safety_checker(image=image_np)
|
329 |
+
for i, not_safe in enumerate(has_nsfw_concept):
|
330 |
+
if not_safe:
|
331 |
+
image[i] = nsfw_image
|
332 |
+
return image
|
requirements.txt
CHANGED
@@ -1,7 +1,7 @@
|
|
1 |
torch==2.1.2
|
2 |
torchvision==0.16.2
|
3 |
accelerate==0.31.0
|
4 |
-
diffusers
|
5 |
huggingface_hub==0.23.4
|
6 |
matplotlib==3.9.1
|
7 |
numpy==1.26.4
|
@@ -12,10 +12,11 @@ scipy==1.13.1
|
|
12 |
setuptools==51.0.0
|
13 |
scikit-image==0.24.0
|
14 |
tqdm==4.66.4
|
15 |
-
transformers==4.
|
16 |
fvcore==0.1.5.post20221221
|
17 |
cloudpickle==3.0.0
|
18 |
omegaconf==2.3.0
|
19 |
pycocotools==2.0.8
|
20 |
av==12.3.0
|
21 |
-
gradio==4.41.0
|
|
|
|
1 |
torch==2.1.2
|
2 |
torchvision==0.16.2
|
3 |
accelerate==0.31.0
|
4 |
+
git+https://github.com/huggingface/diffusers.git
|
5 |
huggingface_hub==0.23.4
|
6 |
matplotlib==3.9.1
|
7 |
numpy==1.26.4
|
|
|
12 |
setuptools==51.0.0
|
13 |
scikit-image==0.24.0
|
14 |
tqdm==4.66.4
|
15 |
+
transformers==4.46.3
|
16 |
fvcore==0.1.5.post20221221
|
17 |
cloudpickle==3.0.0
|
18 |
omegaconf==2.3.0
|
19 |
pycocotools==2.0.8
|
20 |
av==12.3.0
|
21 |
+
gradio==4.41.0
|
22 |
+
peft==0.14.0
|