Upload app.py with huggingface_hub
Browse files
app.py
CHANGED
|
@@ -248,6 +248,47 @@ def estimate_duration(
|
|
| 248 |
|
| 249 |
|
| 250 |
@spaces.GPU(duration=estimate_duration)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 251 |
def generate_image(
|
| 252 |
prompt: str = "",
|
| 253 |
input_images=None,
|
|
@@ -257,7 +298,6 @@ def generate_image(
|
|
| 257 |
seed: int = 0,
|
| 258 |
resolution: int = 1024,
|
| 259 |
aspect_ratio: str = "Auto",
|
| 260 |
-
progress=gr.Progress(track_tqdm=True),
|
| 261 |
):
|
| 262 |
"""Generate or edit an image with Qwen-Image-2.1.
|
| 263 |
|
|
@@ -277,6 +317,9 @@ def generate_image(
|
|
| 277 |
Returns:
|
| 278 |
The generated image, as RGBA (the model natively supports transparency).
|
| 279 |
"""
|
|
|
|
|
|
|
|
|
|
| 280 |
prompt = (prompt or "").strip()
|
| 281 |
if not prompt:
|
| 282 |
raise gr.Error("Please enter a prompt.")
|
|
@@ -285,38 +328,21 @@ def generate_image(
|
|
| 285 |
if len(image_paths) > MAX_INPUT_IMAGES:
|
| 286 |
raise gr.Error(f"Up to {MAX_INPUT_IMAGES} input images are supported.")
|
| 287 |
|
| 288 |
-
#
|
| 289 |
-
#
|
| 290 |
if image_paths:
|
| 291 |
check_guard(prompt)
|
| 292 |
|
| 293 |
-
|
| 294 |
-
|
| 295 |
-
|
| 296 |
-
|
| 297 |
-
|
| 298 |
-
|
| 299 |
-
|
| 300 |
-
|
| 301 |
-
|
| 302 |
-
started = time.perf_counter()
|
| 303 |
-
image = pipe(
|
| 304 |
-
prompt=prompt,
|
| 305 |
-
image=condition_images,
|
| 306 |
-
height=height,
|
| 307 |
-
width=width,
|
| 308 |
-
output_resolution=int(resolution),
|
| 309 |
-
num_inference_steps=int(num_inference_steps),
|
| 310 |
-
generator=torch.Generator(device="cuda").manual_seed(int(seed) % (MAX_SEED + 1)),
|
| 311 |
-
**call_kwargs,
|
| 312 |
-
).images[0]
|
| 313 |
-
print(
|
| 314 |
-
f"[generate] images={len(image_paths)} steps={num_inference_steps} "
|
| 315 |
-
f"res={resolution} ratio={aspect_ratio} size={image.size} "
|
| 316 |
-
f"elapsed={time.perf_counter() - started:.1f}s",
|
| 317 |
-
flush=True,
|
| 318 |
)
|
| 319 |
-
return image
|
| 320 |
|
| 321 |
|
| 322 |
def prepare(prompt, input_images, enhance, seed, randomize_seed, aspect_ratio):
|
|
|
|
| 248 |
|
| 249 |
|
| 250 |
@spaces.GPU(duration=estimate_duration)
|
| 251 |
+
def _diffuse(
|
| 252 |
+
prompt: str,
|
| 253 |
+
image_paths: list,
|
| 254 |
+
negative_prompt: str,
|
| 255 |
+
true_cfg_scale: float,
|
| 256 |
+
num_inference_steps: int,
|
| 257 |
+
seed: int,
|
| 258 |
+
resolution: int,
|
| 259 |
+
aspect_ratio: str,
|
| 260 |
+
progress=gr.Progress(track_tqdm=True),
|
| 261 |
+
):
|
| 262 |
+
"""Run the diffusion pipeline. Inputs are already validated and screened."""
|
| 263 |
+
condition_images = [Image.open(p) for p in image_paths] or None
|
| 264 |
+
width, height = resolve_size(resolution, aspect_ratio)
|
| 265 |
+
|
| 266 |
+
negative_prompt = (negative_prompt or "").strip()
|
| 267 |
+
call_kwargs = {}
|
| 268 |
+
if negative_prompt and float(true_cfg_scale) > 1.0:
|
| 269 |
+
call_kwargs["negative_prompt"] = negative_prompt
|
| 270 |
+
call_kwargs["true_cfg_scale"] = float(true_cfg_scale)
|
| 271 |
+
|
| 272 |
+
started = time.perf_counter()
|
| 273 |
+
image = pipe(
|
| 274 |
+
prompt=prompt,
|
| 275 |
+
image=condition_images,
|
| 276 |
+
height=height,
|
| 277 |
+
width=width,
|
| 278 |
+
output_resolution=int(resolution),
|
| 279 |
+
num_inference_steps=int(num_inference_steps),
|
| 280 |
+
generator=torch.Generator(device="cuda").manual_seed(int(seed) % (MAX_SEED + 1)),
|
| 281 |
+
**call_kwargs,
|
| 282 |
+
).images[0]
|
| 283 |
+
print(
|
| 284 |
+
f"[generate] images={len(image_paths)} steps={num_inference_steps} "
|
| 285 |
+
f"res={resolution} ratio={aspect_ratio} size={image.size} "
|
| 286 |
+
f"elapsed={time.perf_counter() - started:.1f}s",
|
| 287 |
+
flush=True,
|
| 288 |
+
)
|
| 289 |
+
return image
|
| 290 |
+
|
| 291 |
+
|
| 292 |
def generate_image(
|
| 293 |
prompt: str = "",
|
| 294 |
input_images=None,
|
|
|
|
| 298 |
seed: int = 0,
|
| 299 |
resolution: int = 1024,
|
| 300 |
aspect_ratio: str = "Auto",
|
|
|
|
| 301 |
):
|
| 302 |
"""Generate or edit an image with Qwen-Image-2.1.
|
| 303 |
|
|
|
|
| 317 |
Returns:
|
| 318 |
The generated image, as RGBA (the model natively supports transparency).
|
| 319 |
"""
|
| 320 |
+
# Validation and screening deliberately happen here, in the main process: a rejected
|
| 321 |
+
# prompt must never reserve a GPU slot, and gr.Error raised inside the ZeroGPU worker
|
| 322 |
+
# reaches the browser as a bare "RuntimeError" instead of its message.
|
| 323 |
prompt = (prompt or "").strip()
|
| 324 |
if not prompt:
|
| 325 |
raise gr.Error("Please enter a prompt.")
|
|
|
|
| 328 |
if len(image_paths) > MAX_INPUT_IMAGES:
|
| 329 |
raise gr.Error(f"Up to {MAX_INPUT_IMAGES} input images are supported.")
|
| 330 |
|
| 331 |
+
# `prepare` already screened the UI path; this keeps the guard in force for callers
|
| 332 |
+
# that hit /generate directly.
|
| 333 |
if image_paths:
|
| 334 |
check_guard(prompt)
|
| 335 |
|
| 336 |
+
return _diffuse(
|
| 337 |
+
prompt,
|
| 338 |
+
image_paths,
|
| 339 |
+
negative_prompt,
|
| 340 |
+
true_cfg_scale,
|
| 341 |
+
num_inference_steps,
|
| 342 |
+
seed,
|
| 343 |
+
resolution,
|
| 344 |
+
aspect_ratio,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 345 |
)
|
|
|
|
| 346 |
|
| 347 |
|
| 348 |
def prepare(prompt, input_images, enhance, seed, randomize_seed, aspect_ratio):
|