Spaces:
Running on Zero
Running on Zero
Update app.py
Browse files
app.py
CHANGED
|
@@ -1,3 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
| 1 |
import os
|
| 2 |
import gc
|
| 3 |
import gradio as gr
|
|
@@ -5,10 +8,10 @@ import numpy as np
|
|
| 5 |
import random
|
| 6 |
import spaces
|
| 7 |
import torch
|
|
|
|
| 8 |
from diffusers import Flux2KleinPipeline, AutoencoderKLFlux2
|
| 9 |
from PIL import Image
|
| 10 |
from pathlib import Path
|
| 11 |
-
import concurrent.futures
|
| 12 |
import threading
|
| 13 |
from typing import Iterable
|
| 14 |
from gradio.themes import Soft
|
|
@@ -115,9 +118,6 @@ pipe_small_decoder = Flux2KleinPipeline.from_pretrained(
|
|
| 115 |
)
|
| 116 |
pipe_small_decoder.enable_model_cpu_offload()
|
| 117 |
|
| 118 |
-
pipe_lock_standard = threading.Lock()
|
| 119 |
-
pipe_lock_small = threading.Lock()
|
| 120 |
-
|
| 121 |
# ── Helpers ───────────────────────────────────────────────────────────────────
|
| 122 |
|
| 123 |
def update_dimensions_from_image(image_list):
|
|
@@ -168,18 +168,18 @@ def parse_input_images(input_images):
|
|
| 168 |
return None
|
| 169 |
|
| 170 |
|
| 171 |
-
|
| 172 |
-
|
| 173 |
-
|
| 174 |
-
|
| 175 |
-
|
| 176 |
-
|
| 177 |
-
return
|
| 178 |
|
| 179 |
|
| 180 |
# ── Inference ─────────────────────────────────────────────────────────────────
|
| 181 |
|
| 182 |
-
@spaces.GPU(duration=
|
| 183 |
def infer(
|
| 184 |
prompt,
|
| 185 |
input_images=None,
|
|
@@ -212,28 +212,47 @@ def infer(
|
|
| 212 |
if image_list is not None:
|
| 213 |
shared_kwargs["image"] = image_list
|
| 214 |
|
| 215 |
-
|
|
|
|
|
|
|
| 216 |
|
| 217 |
-
|
| 218 |
-
|
| 219 |
-
|
| 220 |
-
|
| 221 |
-
|
| 222 |
-
|
| 223 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 224 |
|
| 225 |
-
|
|
|
|
| 226 |
|
| 227 |
-
|
| 228 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 229 |
|
| 230 |
gc.collect()
|
| 231 |
torch.cuda.empty_cache()
|
| 232 |
|
| 233 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 234 |
|
| 235 |
|
| 236 |
-
@spaces.GPU(duration=
|
| 237 |
def infer_example(images, prompt):
|
| 238 |
if not images:
|
| 239 |
images_list = None
|
|
@@ -241,8 +260,8 @@ def infer_example(images, prompt):
|
|
| 241 |
images_list = [images]
|
| 242 |
else:
|
| 243 |
images_list = images
|
| 244 |
-
|
| 245 |
-
out_std, out_small, seed_used = infer(
|
| 246 |
prompt=prompt,
|
| 247 |
input_images=images_list,
|
| 248 |
seed=0,
|
|
@@ -252,7 +271,7 @@ def infer_example(images, prompt):
|
|
| 252 |
num_inference_steps=4,
|
| 253 |
guidance_scale=1.0,
|
| 254 |
)
|
| 255 |
-
return out_std, out_small, seed_used
|
| 256 |
|
| 257 |
|
| 258 |
# ── CSS ───────────────────────────────────────────────────────────────────────
|
|
@@ -274,6 +293,16 @@ css = """
|
|
| 274 |
display: block;
|
| 275 |
margin-bottom: 6px;
|
| 276 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 277 |
"""
|
| 278 |
|
| 279 |
# ── UI ────────────────────────────────────────────────────────────────────────
|
|
@@ -287,8 +316,8 @@ with gr.Blocks() as demo:
|
|
| 287 |
elem_id="main-title",
|
| 288 |
)
|
| 289 |
gr.Markdown(
|
| 290 |
-
"Compare **FLUX.2-klein-4B** side-by-side with two VAE decoders — generated **
|
| 291 |
-
"🟦 **Standard VAE**
|
| 292 |
)
|
| 293 |
|
| 294 |
# ── Main two-column row ───────────────────────────────────────────────
|
|
@@ -316,11 +345,12 @@ with gr.Blocks() as demo:
|
|
| 316 |
# ── Right: outputs ────────────────────────────────────────────────
|
| 317 |
with gr.Column():
|
| 318 |
with gr.Row():
|
|
|
|
| 319 |
with gr.Column():
|
| 320 |
gr.HTML(
|
| 321 |
'<span class="vae-badge" '
|
| 322 |
'style="background:#FFE0CC;color:#CC3700;">'
|
| 323 |
-
'🟦 Standard VAE</span>'
|
| 324 |
)
|
| 325 |
result_standard = gr.Image(
|
| 326 |
label="Standard VAE",
|
|
@@ -329,12 +359,17 @@ with gr.Blocks() as demo:
|
|
| 329 |
format="png",
|
| 330 |
height=280,
|
| 331 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 332 |
|
|
|
|
| 333 |
with gr.Column():
|
| 334 |
gr.HTML(
|
| 335 |
'<span class="vae-badge" '
|
| 336 |
'style="background:#FFF0E5;color:#E63E00;">'
|
| 337 |
-
'🟩 Small Decoder VAE</span>'
|
| 338 |
)
|
| 339 |
result_small = gr.Image(
|
| 340 |
label="Small Decoder VAE",
|
|
@@ -343,8 +378,12 @@ with gr.Blocks() as demo:
|
|
| 343 |
format="png",
|
| 344 |
height=280,
|
| 345 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
| 346 |
|
| 347 |
-
with gr.Accordion("Advanced Settings", open=False
|
| 348 |
seed_output = gr.Number(label="Seed Used", precision=0)
|
| 349 |
seed = gr.Slider(
|
| 350 |
label="Seed",
|
|
@@ -387,19 +426,20 @@ with gr.Blocks() as demo:
|
|
| 387 |
value=1.0,
|
| 388 |
)
|
| 389 |
|
| 390 |
-
# ──
|
| 391 |
gr.Examples(
|
| 392 |
examples=[
|
| 393 |
[["examples/1.jpg"], "Change the weather to stormy."],
|
| 394 |
[["examples/2.jpg"], "Transform the scene into a snowy winter day while preserving the original subject identity, framing, and composition."],
|
| 395 |
[["examples/3.jpg"], "Relight the image with soft golden sunset lighting while keeping all structures and subject details consistent."],
|
| 396 |
[["examples/4.jpg"], "Make the texture high-resolution."],
|
|
|
|
| 397 |
],
|
| 398 |
inputs=[input_images, prompt],
|
| 399 |
-
outputs=[result_standard, result_small, seed_output],
|
| 400 |
fn=infer_example,
|
| 401 |
cache_examples=False,
|
| 402 |
-
label="Examples"
|
| 403 |
)
|
| 404 |
|
| 405 |
gr.Markdown(
|
|
@@ -428,13 +468,12 @@ with gr.Blocks() as demo:
|
|
| 428 |
num_inference_steps,
|
| 429 |
guidance_scale,
|
| 430 |
],
|
| 431 |
-
outputs=[result_standard, result_small, seed_output],
|
| 432 |
)
|
| 433 |
|
| 434 |
if __name__ == "__main__":
|
| 435 |
demo.queue(max_size=20).launch(
|
| 436 |
-
css=css,
|
| 437 |
-
theme=orange_red_theme,
|
| 438 |
ssr_mode=False,
|
| 439 |
show_error=True,
|
| 440 |
)
|
|
|
|
| 1 |
+
Here's the updated code with sequential generation and individual timing:
|
| 2 |
+
|
| 3 |
+
```python
|
| 4 |
import os
|
| 5 |
import gc
|
| 6 |
import gradio as gr
|
|
|
|
| 8 |
import random
|
| 9 |
import spaces
|
| 10 |
import torch
|
| 11 |
+
import time
|
| 12 |
from diffusers import Flux2KleinPipeline, AutoencoderKLFlux2
|
| 13 |
from PIL import Image
|
| 14 |
from pathlib import Path
|
|
|
|
| 15 |
import threading
|
| 16 |
from typing import Iterable
|
| 17 |
from gradio.themes import Soft
|
|
|
|
| 118 |
)
|
| 119 |
pipe_small_decoder.enable_model_cpu_offload()
|
| 120 |
|
|
|
|
|
|
|
|
|
|
| 121 |
# ── Helpers ───────────────────────────────────────────────────────────────────
|
| 122 |
|
| 123 |
def update_dimensions_from_image(image_list):
|
|
|
|
| 168 |
return None
|
| 169 |
|
| 170 |
|
| 171 |
+
def format_time(seconds: float) -> str:
|
| 172 |
+
"""Format seconds into a human-readable string."""
|
| 173 |
+
if seconds < 60:
|
| 174 |
+
return f"{seconds:.2f}s"
|
| 175 |
+
minutes = int(seconds // 60)
|
| 176 |
+
secs = seconds % 60
|
| 177 |
+
return f"{minutes}m {secs:.2f}s"
|
| 178 |
|
| 179 |
|
| 180 |
# ── Inference ─────────────────────────────────────────────────────────────────
|
| 181 |
|
| 182 |
+
@spaces.GPU(duration=240)
|
| 183 |
def infer(
|
| 184 |
prompt,
|
| 185 |
input_images=None,
|
|
|
|
| 212 |
if image_list is not None:
|
| 213 |
shared_kwargs["image"] = image_list
|
| 214 |
|
| 215 |
+
# ── Step 1: Standard VAE ──────────────────────────────────────────────────
|
| 216 |
+
progress(0.05, desc="🟦 Running Standard VAE generation...")
|
| 217 |
+
print("Starting Standard VAE generation...")
|
| 218 |
|
| 219 |
+
t0_std = time.perf_counter()
|
| 220 |
+
gen_std = torch.Generator(device="cpu").manual_seed(seed)
|
| 221 |
+
out_standard = pipe_standard(**shared_kwargs, generator=gen_std).images[0]
|
| 222 |
+
t1_std = time.perf_counter()
|
| 223 |
+
time_std = t1_std - t0_std
|
| 224 |
+
time_std_str = format_time(time_std)
|
| 225 |
+
|
| 226 |
+
print(f"Standard VAE done in {time_std_str}")
|
| 227 |
+
progress(0.55, desc=f"🟦 Standard VAE done in {time_std_str} — now running 🟩 Small Decoder VAE...")
|
| 228 |
+
|
| 229 |
+
gc.collect()
|
| 230 |
+
torch.cuda.empty_cache()
|
| 231 |
|
| 232 |
+
# ── Step 2: Small Decoder VAE ─────────────────────────────────────────────
|
| 233 |
+
print("Starting Small Decoder VAE generation...")
|
| 234 |
|
| 235 |
+
t0_small = time.perf_counter()
|
| 236 |
+
gen_small = torch.Generator(device="cpu").manual_seed(seed)
|
| 237 |
+
out_small = pipe_small_decoder(**shared_kwargs, generator=gen_small).images[0]
|
| 238 |
+
t1_small = time.perf_counter()
|
| 239 |
+
time_small = t1_small - t0_small
|
| 240 |
+
time_small_str = format_time(time_small)
|
| 241 |
+
|
| 242 |
+
print(f"Small Decoder VAE done in {time_small_str}")
|
| 243 |
+
progress(1.0, desc=f"✅ Both done! Standard: {time_std_str} | Small Decoder: {time_small_str}")
|
| 244 |
|
| 245 |
gc.collect()
|
| 246 |
torch.cuda.empty_cache()
|
| 247 |
|
| 248 |
+
# ── Build timing label strings ────────────────────────────────────────────
|
| 249 |
+
label_std = f"🟦 Standard VAE — ⏱ {time_std_str}"
|
| 250 |
+
label_small = f"🟩 Small Decoder VAE — ⏱ {time_small_str}"
|
| 251 |
+
|
| 252 |
+
return out_standard, out_small, seed, label_std, label_small
|
| 253 |
|
| 254 |
|
| 255 |
+
@spaces.GPU(duration=240)
|
| 256 |
def infer_example(images, prompt):
|
| 257 |
if not images:
|
| 258 |
images_list = None
|
|
|
|
| 260 |
images_list = [images]
|
| 261 |
else:
|
| 262 |
images_list = images
|
| 263 |
+
|
| 264 |
+
out_std, out_small, seed_used, label_std, label_small = infer(
|
| 265 |
prompt=prompt,
|
| 266 |
input_images=images_list,
|
| 267 |
seed=0,
|
|
|
|
| 271 |
num_inference_steps=4,
|
| 272 |
guidance_scale=1.0,
|
| 273 |
)
|
| 274 |
+
return out_std, out_small, seed_used, label_std, label_small
|
| 275 |
|
| 276 |
|
| 277 |
# ── CSS ───────────────────────────────────────────────────────────────────────
|
|
|
|
| 293 |
display: block;
|
| 294 |
margin-bottom: 6px;
|
| 295 |
}
|
| 296 |
+
.timing-label {
|
| 297 |
+
text-align: center;
|
| 298 |
+
font-size: 0.92em;
|
| 299 |
+
font-weight: 600;
|
| 300 |
+
color: #555;
|
| 301 |
+
margin-top: 4px;
|
| 302 |
+
padding: 3px 10px;
|
| 303 |
+
border-radius: 12px;
|
| 304 |
+
background: rgba(255,255,255,0.6);
|
| 305 |
+
}
|
| 306 |
"""
|
| 307 |
|
| 308 |
# ── UI ────────────────────────────────────────────────────────────────────────
|
|
|
|
| 316 |
elem_id="main-title",
|
| 317 |
)
|
| 318 |
gr.Markdown(
|
| 319 |
+
"Compare **FLUX.2-klein-4B** side-by-side with two VAE decoders — generated **sequentially** from the **same seed**. \n"
|
| 320 |
+
"🟦 **Standard VAE** (generated first) → 🟩 **Small Decoder VAE** ([FLUX.2-small-decoder](https://huggingface.co/black-forest-labs/FLUX.2-small-decoder)) · [[model](https://huggingface.co/black-forest-labs/FLUX.2-klein-4B)]"
|
| 321 |
)
|
| 322 |
|
| 323 |
# ── Main two-column row ───────────────────────────────────────────────
|
|
|
|
| 345 |
# ── Right: outputs ────────────────────────────────────────────────
|
| 346 |
with gr.Column():
|
| 347 |
with gr.Row():
|
| 348 |
+
# Standard VAE output
|
| 349 |
with gr.Column():
|
| 350 |
gr.HTML(
|
| 351 |
'<span class="vae-badge" '
|
| 352 |
'style="background:#FFE0CC;color:#CC3700;">'
|
| 353 |
+
'🟦 Standard VAE · Generated First</span>'
|
| 354 |
)
|
| 355 |
result_standard = gr.Image(
|
| 356 |
label="Standard VAE",
|
|
|
|
| 359 |
format="png",
|
| 360 |
height=280,
|
| 361 |
)
|
| 362 |
+
timing_standard = gr.Markdown(
|
| 363 |
+
value="⏱ Waiting...",
|
| 364 |
+
elem_classes=["timing-label"],
|
| 365 |
+
)
|
| 366 |
|
| 367 |
+
# Small Decoder VAE output
|
| 368 |
with gr.Column():
|
| 369 |
gr.HTML(
|
| 370 |
'<span class="vae-badge" '
|
| 371 |
'style="background:#FFF0E5;color:#E63E00;">'
|
| 372 |
+
'🟩 Small Decoder VAE · Generated Second</span>'
|
| 373 |
)
|
| 374 |
result_small = gr.Image(
|
| 375 |
label="Small Decoder VAE",
|
|
|
|
| 378 |
format="png",
|
| 379 |
height=280,
|
| 380 |
)
|
| 381 |
+
timing_small = gr.Markdown(
|
| 382 |
+
value="⏱ Waiting...",
|
| 383 |
+
elem_classes=["timing-label"],
|
| 384 |
+
)
|
| 385 |
|
| 386 |
+
with gr.Accordion("Advanced Settings", open=False):
|
| 387 |
seed_output = gr.Number(label="Seed Used", precision=0)
|
| 388 |
seed = gr.Slider(
|
| 389 |
label="Seed",
|
|
|
|
| 426 |
value=1.0,
|
| 427 |
)
|
| 428 |
|
| 429 |
+
# ── Examples ──────────────────────────────────────────────────────────
|
| 430 |
gr.Examples(
|
| 431 |
examples=[
|
| 432 |
[["examples/1.jpg"], "Change the weather to stormy."],
|
| 433 |
[["examples/2.jpg"], "Transform the scene into a snowy winter day while preserving the original subject identity, framing, and composition."],
|
| 434 |
[["examples/3.jpg"], "Relight the image with soft golden sunset lighting while keeping all structures and subject details consistent."],
|
| 435 |
[["examples/4.jpg"], "Make the texture high-resolution."],
|
| 436 |
+
[None, "A beautiful cyberpunk cityscape at night, neon lights, highly detailed."],
|
| 437 |
],
|
| 438 |
inputs=[input_images, prompt],
|
| 439 |
+
outputs=[result_standard, result_small, seed_output, timing_standard, timing_small],
|
| 440 |
fn=infer_example,
|
| 441 |
cache_examples=False,
|
| 442 |
+
label="Examples",
|
| 443 |
)
|
| 444 |
|
| 445 |
gr.Markdown(
|
|
|
|
| 468 |
num_inference_steps,
|
| 469 |
guidance_scale,
|
| 470 |
],
|
| 471 |
+
outputs=[result_standard, result_small, seed_output, timing_standard, timing_small],
|
| 472 |
)
|
| 473 |
|
| 474 |
if __name__ == "__main__":
|
| 475 |
demo.queue(max_size=20).launch(
|
| 476 |
+
css=css, theme=orange_red_theme,
|
|
|
|
| 477 |
ssr_mode=False,
|
| 478 |
show_error=True,
|
| 479 |
)
|