Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -13,15 +13,17 @@ import gradio as gr
|
|
13 |
import spaces
|
14 |
from diffusers import (
|
15 |
DiffusionPipeline,
|
16 |
-
FlowMatchEulerDiscreteScheduler,
|
17 |
-
AutoencoderTiny,
|
18 |
AutoencoderKL,
|
19 |
-
|
|
|
|
|
|
|
20 |
from huggingface_hub import (
|
21 |
hf_hub_download,
|
22 |
HfFileSystem,
|
23 |
ModelCard,
|
24 |
-
snapshot_download
|
|
|
25 |
from diffusers.utils import load_image
|
26 |
import requests
|
27 |
from urllib.parse import urlparse
|
@@ -118,10 +120,14 @@ loras = [
|
|
118 |
},
|
119 |
]
|
120 |
|
121 |
-
# Initialize the base model
|
122 |
dtype = torch.bfloat16
|
123 |
base_model = "Qwen/Qwen-Image"
|
124 |
|
|
|
|
|
|
|
|
|
125 |
# Scheduler configuration from the Qwen-Image-Lightning repository
|
126 |
scheduler_config = {
|
127 |
"base_image_seq_len": 256,
|
@@ -141,34 +147,30 @@ scheduler_config = {
|
|
141 |
}
|
142 |
|
143 |
scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)
|
|
|
|
|
144 |
pipe = DiffusionPipeline.from_pretrained(
|
145 |
-
base_model, scheduler=scheduler, torch_dtype=dtype
|
146 |
).to(device)
|
147 |
|
148 |
-
|
149 |
-
good_vae = AutoencoderKL.from_pretrained(base_model, subfolder="vae", torch_dtype=dtype).to(device)
|
150 |
-
pipe.vae = taef1
|
151 |
pipe_i2i = AutoPipelineForImage2Image.from_pretrained(
|
152 |
base_model,
|
153 |
vae=good_vae,
|
154 |
-
transformer=pipe.transformer,
|
155 |
-
text_encoder=pipe.text_encoder,
|
156 |
-
tokenizer=pipe.tokenizer,
|
157 |
-
text_encoder_2=pipe.text_encoder_2,
|
158 |
-
tokenizer_2=pipe.tokenizer_2,
|
159 |
scheduler=scheduler,
|
160 |
torch_dtype=dtype
|
161 |
).to(device)
|
162 |
|
|
|
163 |
# Lightning LoRA info (no global state)
|
164 |
LIGHTNING_LORA_REPO = "lightx2v/Qwen-Image-Lightning"
|
165 |
LIGHTNING_LORA_WEIGHT = "Qwen-Image-Lightning-8steps-V1.0.safetensors"
|
166 |
|
167 |
-
MAX_SEED =
|
168 |
|
169 |
-
class
|
170 |
-
def __init__(self,
|
171 |
-
self.
|
172 |
|
173 |
def __enter__(self):
|
174 |
self.start_time = time.time()
|
@@ -177,8 +179,8 @@ class calculateDuration:
|
|
177 |
def __exit__(self, exc_type, exc_value, traceback):
|
178 |
self.end_time = time.time()
|
179 |
self.elapsed_time = self.end_time - self.start_time
|
180 |
-
if self.
|
181 |
-
print(f"Elapsed time for {self.
|
182 |
else:
|
183 |
print(f"Elapsed time: {self.elapsed_time:.6f} seconds")
|
184 |
|
@@ -230,88 +232,32 @@ def adjust_generation_mode(speed_mode):
|
|
230 |
else:
|
231 |
return gr.update(value="Base mode selected - 48 steps for best quality"), 48, 4.0
|
232 |
|
233 |
-
|
234 |
-
|
235 |
-
generator = torch.Generator(device="cuda").manual_seed(seed)
|
236 |
-
pipe.to("cuda")
|
237 |
-
|
238 |
-
batch_size = 1
|
239 |
-
prompt = prompt_mash
|
240 |
-
do_classifier_free_guidance = cfg_scale > 1.0
|
241 |
-
prompt_embeds, pooled_prompt_embeds = pipe.encode_prompt(
|
242 |
-
prompt,
|
243 |
-
num_images_per_prompt=1,
|
244 |
-
do_classifier_free_guidance=do_classifier_free_guidance,
|
245 |
-
prompt_2=None,
|
246 |
-
max_sequence_length=256,
|
247 |
-
)
|
248 |
-
height, width = height - height % 16, width - width % 16
|
249 |
-
latents = pipe.prepare_latents(
|
250 |
-
batch_size,
|
251 |
-
pipe.transformer.config.in_channels,
|
252 |
-
height,
|
253 |
-
width,
|
254 |
-
dtype,
|
255 |
-
device,
|
256 |
-
generator,
|
257 |
-
latents=None,
|
258 |
-
)
|
259 |
-
pipe.scheduler.set_timesteps(steps)
|
260 |
-
timesteps = pipe.scheduler.timesteps
|
261 |
-
joint_attention_kwargs = {"scale": lora_scale}
|
262 |
-
for i in range(steps):
|
263 |
-
t = pipe.scheduler.sigmas[i]
|
264 |
-
latent_model_input = latents
|
265 |
-
with torch.no_grad():
|
266 |
-
noise_pred = pipe.transformer(
|
267 |
-
hidden_states=latent_model_input,
|
268 |
-
timestep=t,
|
269 |
-
guidance=cfg_scale,
|
270 |
-
pooled_projections=pooled_prompt_embeds,
|
271 |
-
encoder_hidden_states=prompt_embeds,
|
272 |
-
joint_attention_kwargs=joint_attention_kwargs,
|
273 |
-
return_dict=False,
|
274 |
-
)[0]
|
275 |
-
latents = pipe.scheduler.step(
|
276 |
-
model_output=noise_pred,
|
277 |
-
timestep=t,
|
278 |
-
sample=latent_model_input,
|
279 |
-
return_dict=False,
|
280 |
-
)[0]
|
281 |
-
# preview
|
282 |
-
with torch.no_grad():
|
283 |
-
decoded = pipe.vae.decode(latents / pipe.vae.config.scaling_factor, return_dict=False)[0]
|
284 |
-
image = pipe.image_processor.pt_to_pil(decoded)[0]
|
285 |
-
yield image
|
286 |
-
# final
|
287 |
-
with torch.no_grad():
|
288 |
-
decoded = good_vae.decode(latents / good_vae.config.scaling_factor, return_dict=False)[0]
|
289 |
-
image = pipe.image_processor.pt_to_pil(decoded)[0]
|
290 |
-
yield image
|
291 |
-
|
292 |
-
@spaces.GPU(duration=100)
|
293 |
-
def generate_image_to_image(prompt_mash, image_input_path, image_strength, steps, cfg_scale, width, height, lora_scale, seed):
|
294 |
generator = torch.Generator(device="cuda").manual_seed(seed)
|
295 |
pipe_i2i.to("cuda")
|
296 |
-
|
|
|
|
|
|
|
297 |
final_image = pipe_i2i(
|
298 |
prompt=prompt_mash,
|
299 |
-
image=
|
300 |
-
strength=
|
301 |
num_inference_steps=steps,
|
302 |
guidance_scale=cfg_scale,
|
303 |
-
width=width,
|
304 |
-
height=height,
|
305 |
generator=generator,
|
306 |
-
|
307 |
-
output_type="pil",
|
308 |
).images[0]
|
309 |
return final_image
|
310 |
|
311 |
@spaces.GPU(duration=100)
|
312 |
-
def
|
|
|
|
|
|
|
313 |
if selected_index is None:
|
314 |
-
raise gr.Error("You must select a LoRA before proceeding
|
315 |
|
316 |
selected_lora = loras[selected_index]
|
317 |
lora_path = selected_lora["repo"]
|
@@ -319,72 +265,85 @@ def process_adapter_generation(prompt, cfg_scale, steps, selected_index, randomi
|
|
319 |
|
320 |
# Prepare prompt with trigger word
|
321 |
if trigger_word:
|
322 |
-
|
323 |
-
if selected_lora["trigger_position"] == "prepend":
|
324 |
-
prompt_mash = f"{trigger_word} {prompt}"
|
325 |
-
else:
|
326 |
-
prompt_mash = f"{prompt} {trigger_word}"
|
327 |
-
else:
|
328 |
-
prompt_mash = f"{trigger_word} {prompt}"
|
329 |
else:
|
330 |
prompt_mash = prompt
|
331 |
|
332 |
-
#
|
333 |
-
|
334 |
-
|
335 |
-
|
336 |
-
|
337 |
pipe_to_use = pipe_i2i if image_input is not None else pipe
|
|
|
|
|
|
|
|
|
338 |
|
|
|
339 |
if speed_mode == "Fast (8 steps)":
|
340 |
-
with
|
341 |
-
# Load Lightning LoRA first
|
342 |
pipe_to_use.load_lora_weights(
|
343 |
LIGHTNING_LORA_REPO,
|
344 |
weight_name=LIGHTNING_LORA_WEIGHT,
|
345 |
adapter_name="lightning"
|
346 |
)
|
347 |
-
|
348 |
-
# Load the selected style LoRA
|
349 |
-
weight_name = selected_lora.get("weights", None)
|
350 |
pipe_to_use.load_lora_weights(
|
351 |
lora_path,
|
352 |
weight_name=weight_name,
|
353 |
-
low_cpu_mem_usage=True,
|
354 |
adapter_name="style"
|
355 |
)
|
356 |
-
|
357 |
-
# Set both adapters active with their weights
|
358 |
pipe_to_use.set_adapters(["lightning", "style"], adapter_weights=[1.0, lora_scale])
|
359 |
-
else:
|
360 |
-
|
361 |
-
|
362 |
-
|
363 |
-
pipe_to_use.load_lora_weights(
|
364 |
-
lora_path,
|
365 |
-
weight_name=weight_name,
|
366 |
-
low_cpu_mem_usage=True
|
367 |
-
)
|
368 |
-
|
369 |
-
# Set random seed for reproducibility
|
370 |
-
with calculateDuration("Randomizing seed"):
|
371 |
-
if randomize_seed:
|
372 |
-
seed = random.randint(0, MAX_SEED)
|
373 |
|
374 |
-
# Get image dimensions from aspect ratio
|
375 |
width, height = compute_image_dimensions(aspect_ratio)
|
376 |
-
|
|
|
377 |
if image_input is not None:
|
378 |
-
|
|
|
379 |
yield final_image, seed, gr.update(visible=False)
|
380 |
else:
|
381 |
-
|
382 |
-
|
383 |
-
|
384 |
-
|
385 |
-
|
386 |
-
|
387 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
388 |
|
389 |
def fetch_hf_adapter_files(link):
|
390 |
split_link = link.split("/")
|
@@ -393,79 +352,37 @@ def fetch_hf_adapter_files(link):
|
|
393 |
|
394 |
print(f"Repository attempted: {split_link}")
|
395 |
|
396 |
-
# Load model card
|
397 |
model_card = ModelCard.load(link)
|
398 |
base_model = model_card.data.get("base_model")
|
399 |
print(f"Base model: {base_model}")
|
400 |
|
401 |
-
# Validate model type (for Qwen-Image)
|
402 |
acceptable_models = {"Qwen/Qwen-Image"}
|
403 |
-
|
404 |
models_to_check = base_model if isinstance(base_model, list) else [base_model]
|
405 |
|
406 |
if not any(model in acceptable_models for model in models_to_check):
|
407 |
raise Exception("Not a Qwen-Image LoRA!")
|
408 |
|
409 |
-
|
410 |
-
image_path = model_card.data.get("widget", [{}])[0].get("output", {}).get("url", None)
|
411 |
trigger_word = model_card.data.get("instance_prompt", "")
|
412 |
image_url = f"https://huggingface.co/{link}/resolve/main/{image_path}" if image_path else None
|
413 |
|
414 |
-
# Initialize Hugging Face file system
|
415 |
fs = HfFileSystem()
|
416 |
try:
|
417 |
list_of_files = fs.ls(link, detail=False)
|
418 |
-
|
419 |
-
# Find safetensors file
|
420 |
-
safetensors_name = None
|
421 |
-
for file in list_of_files:
|
422 |
-
filename = file.split("/")[-1]
|
423 |
-
if filename.endswith(".safetensors"):
|
424 |
-
safetensors_name = filename
|
425 |
-
break
|
426 |
-
|
427 |
if not safetensors_name:
|
428 |
raise Exception("No valid *.safetensors file found in the repository.")
|
429 |
-
|
430 |
except Exception as e:
|
431 |
print(e)
|
432 |
-
raise Exception("
|
433 |
|
434 |
return split_link[1], link, safetensors_name, trigger_word, image_url
|
435 |
|
436 |
def validate_custom_adapter(link):
|
437 |
print(f"Checking a custom model on: {link}")
|
438 |
-
|
439 |
-
|
440 |
-
|
441 |
-
parts = link.split('/')
|
442 |
-
try:
|
443 |
-
hf_index = parts.index('huggingface.co')
|
444 |
-
username = parts[hf_index + 1]
|
445 |
-
repo_name = parts[hf_index + 2]
|
446 |
-
repo = f"{username}/{repo_name}"
|
447 |
-
|
448 |
-
safetensors_name = parts[-1]
|
449 |
-
|
450 |
-
try:
|
451 |
-
model_card = ModelCard.load(repo)
|
452 |
-
trigger_word = model_card.data.get("instance_prompt", "")
|
453 |
-
image_path = model_card.data.get("widget", [{}])[0].get("output", {}).get("url", None)
|
454 |
-
image_url = f"https://huggingface.co/{repo}/resolve/main/{image_path}" if image_path else None
|
455 |
-
except:
|
456 |
-
trigger_word = ""
|
457 |
-
image_url = None
|
458 |
-
|
459 |
-
return repo_name, repo, safetensors_name, trigger_word, image_url
|
460 |
-
except:
|
461 |
-
raise Exception("Invalid safetensors URL format")
|
462 |
-
|
463 |
-
if link.startswith("https://"):
|
464 |
-
if link.startswith("https://huggingface.co") or link.startswith("https://www.huggingface.co"):
|
465 |
-
link_split = link.split("huggingface.co/")
|
466 |
-
return fetch_hf_adapter_files(link_split[1])
|
467 |
-
else:
|
468 |
-
return fetch_hf_adapter_files(link)
|
469 |
|
470 |
def incorporate_custom_adapter(custom_lora):
|
471 |
global loras
|
@@ -485,29 +402,22 @@ def incorporate_custom_adapter(custom_lora):
|
|
485 |
</div>
|
486 |
</div>
|
487 |
'''
|
488 |
-
existing_item_index = next((
|
489 |
if existing_item_index is None:
|
490 |
-
new_item = {
|
491 |
-
"image": image,
|
492 |
-
"title": title,
|
493 |
-
"repo": repo,
|
494 |
-
"weights": path,
|
495 |
-
"trigger_word": trigger_word
|
496 |
-
}
|
497 |
-
print(new_item)
|
498 |
loras.append(new_item)
|
499 |
-
existing_item_index = len(loras) - 1
|
500 |
|
501 |
return gr.update(visible=True, value=card), gr.update(visible=True), gr.Gallery(selected_index=None), f"Custom: {path}", existing_item_index, trigger_word
|
502 |
except Exception as e:
|
503 |
-
gr.Warning(f"Invalid LoRA:
|
504 |
-
return gr.update(visible=True, value=f"Invalid LoRA:
|
505 |
-
|
506 |
-
return gr.update(visible=False), gr.update(visible=False), gr.update(), "", None, ""
|
507 |
|
508 |
def discard_custom_adapter():
|
509 |
return gr.update(visible=False), gr.update(visible=False), gr.update(), "", None, ""
|
510 |
|
|
|
511 |
css = '''
|
512 |
#gen_btn{height: 100%}
|
513 |
#gen_column{align-self: stretch}
|
@@ -523,7 +433,7 @@ css = '''
|
|
523 |
#progress{height:30px}
|
524 |
#progress .generating{display:none}
|
525 |
.progress-container {width: 100%;height: 30px;background-color: #f0f0f0;border-radius: 15px;overflow: hidden;margin-bottom: 20px}
|
526 |
-
.progress-bar {height: 100%;background-color: #4f46e5;width: calc(var(--current) / var(--total) * 100%);transition: width 0.
|
527 |
'''
|
528 |
|
529 |
with gr.Blocks(theme="bethecloud/storj_theme", css=css, delete_cache=(120, 120)) as app:
|
@@ -547,6 +457,10 @@ with gr.Blocks(theme="bethecloud/storj_theme", css=css, delete_cache=(120, 120))
|
|
547 |
elem_id="gallery",
|
548 |
show_share_button=False
|
549 |
)
|
|
|
|
|
|
|
|
|
550 |
with gr.Group():
|
551 |
custom_lora = gr.Textbox(label="Custom LoRA", info="LoRA Hugging Face path", placeholder="username/lora-model-name")
|
552 |
gr.Markdown("[Check Qwen-Image LoRAs](https://huggingface.co/models?other=base_model:adapter:Qwen/Qwen-Image)", elem_id="lora_list")
|
@@ -555,14 +469,14 @@ with gr.Blocks(theme="bethecloud/storj_theme", css=css, delete_cache=(120, 120))
|
|
555 |
|
556 |
with gr.Column():
|
557 |
result = gr.Image(label="Generated Image")
|
558 |
-
|
559 |
|
560 |
with gr.Row():
|
561 |
aspect_ratio = gr.Dropdown(
|
562 |
label="Aspect Ratio",
|
563 |
choices=["1:1", "16:9", "9:16", "4:3", "3:4", "3:2", "2:3"],
|
564 |
value="1:1"
|
565 |
-
|
566 |
with gr.Row():
|
567 |
speed_mode = gr.Dropdown(
|
568 |
label="Output Mode",
|
@@ -577,12 +491,12 @@ with gr.Blocks(theme="bethecloud/storj_theme", css=css, delete_cache=(120, 120))
|
|
577 |
with gr.Column():
|
578 |
with gr.Row():
|
579 |
cfg_scale = gr.Slider(
|
580 |
-
label="Guidance Scale
|
581 |
minimum=1.0,
|
582 |
maximum=5.0,
|
583 |
step=0.1,
|
584 |
value=4.0,
|
585 |
-
info="Lower for speed mode, higher for quality"
|
586 |
)
|
587 |
steps = gr.Slider(
|
588 |
label="Steps",
|
@@ -597,10 +511,6 @@ with gr.Blocks(theme="bethecloud/storj_theme", css=css, delete_cache=(120, 120))
|
|
597 |
randomize_seed = gr.Checkbox(True, label="Randomize seed")
|
598 |
seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0, randomize=True)
|
599 |
lora_scale = gr.Slider(label="LoRA Scale", minimum=0, maximum=2, step=0.01, value=1.0)
|
600 |
-
|
601 |
-
with gr.Row():
|
602 |
-
image_input = gr.Image(label="Input Image for Image2Image", type="filepath")
|
603 |
-
image_strength = gr.Slider(label="Image Strength", minimum=0, maximum=1, step=0.01, value=0.35)
|
604 |
|
605 |
# Event handlers
|
606 |
gallery.select(
|
@@ -626,11 +536,18 @@ with gr.Blocks(theme="bethecloud/storj_theme", css=css, delete_cache=(120, 120))
|
|
626 |
outputs=[custom_lora_info, custom_lora_button, gallery, selected_info, selected_index, custom_lora]
|
627 |
)
|
628 |
|
629 |
-
|
630 |
-
|
631 |
-
|
632 |
-
|
633 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
634 |
)
|
635 |
|
636 |
app.queue()
|
|
|
13 |
import spaces
|
14 |
from diffusers import (
|
15 |
DiffusionPipeline,
|
|
|
|
|
16 |
AutoencoderKL,
|
17 |
+
AutoencoderTiny,
|
18 |
+
AutoPipelineForImage2Image,
|
19 |
+
FlowMatchEulerDiscreteScheduler
|
20 |
+
)
|
21 |
from huggingface_hub import (
|
22 |
hf_hub_download,
|
23 |
HfFileSystem,
|
24 |
ModelCard,
|
25 |
+
snapshot_download
|
26 |
+
)
|
27 |
from diffusers.utils import load_image
|
28 |
import requests
|
29 |
from urllib.parse import urlparse
|
|
|
120 |
},
|
121 |
]
|
122 |
|
123 |
+
# Initialize the base model and autoencoders
|
124 |
dtype = torch.bfloat16
|
125 |
base_model = "Qwen/Qwen-Image"
|
126 |
|
127 |
+
# Initialize TAEF1 for fast previews and the standard VAE for high-quality final images
|
128 |
+
taef1 = AutoencoderTiny.from_pretrained("madebyollin/taef1", torch_dtype=dtype).to(device)
|
129 |
+
good_vae = AutoencoderKL.from_pretrained(base_model, subfolder="vae", torch_dtype=dtype).to(device)
|
130 |
+
|
131 |
# Scheduler configuration from the Qwen-Image-Lightning repository
|
132 |
scheduler_config = {
|
133 |
"base_image_seq_len": 256,
|
|
|
147 |
}
|
148 |
|
149 |
scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)
|
150 |
+
|
151 |
+
# Main pipeline for text-to-image, using taef1 for fast decoding during generation
|
152 |
pipe = DiffusionPipeline.from_pretrained(
|
153 |
+
base_model, scheduler=scheduler, torch_dtype=dtype, vae=taef1
|
154 |
).to(device)
|
155 |
|
156 |
+
# Image-to-image pipeline, using the high-quality VAE
|
|
|
|
|
157 |
pipe_i2i = AutoPipelineForImage2Image.from_pretrained(
|
158 |
base_model,
|
159 |
vae=good_vae,
|
|
|
|
|
|
|
|
|
|
|
160 |
scheduler=scheduler,
|
161 |
torch_dtype=dtype
|
162 |
).to(device)
|
163 |
|
164 |
+
|
165 |
# Lightning LoRA info (no global state)
|
166 |
LIGHTNING_LORA_REPO = "lightx2v/Qwen-Image-Lightning"
|
167 |
LIGHTNING_LORA_WEIGHT = "Qwen-Image-Lightning-8steps-V1.0.safetensors"
|
168 |
|
169 |
+
MAX_SEED = np.iinfo(np.int32).max
|
170 |
|
171 |
+
class Timer:
|
172 |
+
def __init__(self, task_name=""):
|
173 |
+
self.task_name = task_name
|
174 |
|
175 |
def __enter__(self):
|
176 |
self.start_time = time.time()
|
|
|
179 |
def __exit__(self, exc_type, exc_value, traceback):
|
180 |
self.end_time = time.time()
|
181 |
self.elapsed_time = self.end_time - self.start_time
|
182 |
+
if self.task_name:
|
183 |
+
print(f"Elapsed time for {self.task_name}: {self.elapsed_time:.6f} seconds")
|
184 |
else:
|
185 |
print(f"Elapsed time: {self.elapsed_time:.6f} seconds")
|
186 |
|
|
|
232 |
else:
|
233 |
return gr.update(value="Base mode selected - 48 steps for best quality"), 48, 4.0
|
234 |
|
235 |
+
def image_to_image_generation(prompt_mash, image_input, strength, steps, cfg_scale, width, height, lora_scale, seed):
|
236 |
+
"""Handles the image-to-image generation process."""
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
237 |
generator = torch.Generator(device="cuda").manual_seed(seed)
|
238 |
pipe_i2i.to("cuda")
|
239 |
+
|
240 |
+
# Resize and convert input image
|
241 |
+
image_input_pil = load_image(image_input).resize((width, height), Image.Resampling.LANCZOS)
|
242 |
+
|
243 |
final_image = pipe_i2i(
|
244 |
prompt=prompt_mash,
|
245 |
+
image=image_input_pil,
|
246 |
+
strength=strength,
|
247 |
num_inference_steps=steps,
|
248 |
guidance_scale=cfg_scale,
|
|
|
|
|
249 |
generator=generator,
|
250 |
+
# Note: image-to-image with Qwen doesn't use `true_cfg_scale`
|
|
|
251 |
).images[0]
|
252 |
return final_image
|
253 |
|
254 |
@spaces.GPU(duration=100)
|
255 |
+
def process_generation_request(
|
256 |
+
prompt, image_input, image_strength, cfg_scale, steps, selected_index,
|
257 |
+
randomize_seed, seed, aspect_ratio, lora_scale, speed_mode, progress=gr.Progress(track_tqdm=True)
|
258 |
+
):
|
259 |
if selected_index is None:
|
260 |
+
raise gr.Error("You must select a LoRA before proceeding.🧨")
|
261 |
|
262 |
selected_lora = loras[selected_index]
|
263 |
lora_path = selected_lora["repo"]
|
|
|
265 |
|
266 |
# Prepare prompt with trigger word
|
267 |
if trigger_word:
|
268 |
+
prompt_mash = f"{trigger_word}, {prompt}" if prompt else trigger_word
|
|
|
|
|
|
|
|
|
|
|
|
|
269 |
else:
|
270 |
prompt_mash = prompt
|
271 |
|
272 |
+
# Set random seed if requested
|
273 |
+
if randomize_seed:
|
274 |
+
seed = random.randint(0, MAX_SEED)
|
275 |
+
|
276 |
+
# Determine which pipeline to use
|
277 |
pipe_to_use = pipe_i2i if image_input is not None else pipe
|
278 |
+
|
279 |
+
# Always unload any existing LoRAs first to avoid conflicts
|
280 |
+
with Timer("Unloading existing LoRAs"):
|
281 |
+
pipe_to_use.unload_lora_weights()
|
282 |
|
283 |
+
# Load LoRAs based on speed mode
|
284 |
if speed_mode == "Fast (8 steps)":
|
285 |
+
with Timer("Loading Lightning LoRA and style LoRA"):
|
|
|
286 |
pipe_to_use.load_lora_weights(
|
287 |
LIGHTNING_LORA_REPO,
|
288 |
weight_name=LIGHTNING_LORA_WEIGHT,
|
289 |
adapter_name="lightning"
|
290 |
)
|
291 |
+
weight_name = selected_lora.get("weights")
|
|
|
|
|
292 |
pipe_to_use.load_lora_weights(
|
293 |
lora_path,
|
294 |
weight_name=weight_name,
|
|
|
295 |
adapter_name="style"
|
296 |
)
|
|
|
|
|
297 |
pipe_to_use.set_adapters(["lightning", "style"], adapter_weights=[1.0, lora_scale])
|
298 |
+
else: # Quality mode
|
299 |
+
with Timer(f"Loading LoRA weights for {selected_lora['title']}"):
|
300 |
+
weight_name = selected_lora.get("weights")
|
301 |
+
pipe_to_use.load_lora_weights(lora_path, weight_name=weight_name)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
302 |
|
|
|
303 |
width, height = compute_image_dimensions(aspect_ratio)
|
304 |
+
|
305 |
+
# --- Generation ---
|
306 |
if image_input is not None:
|
307 |
+
# Image-to-Image Generation
|
308 |
+
final_image = image_to_image_generation(prompt_mash, image_input, image_strength, steps, cfg_scale, width, height, lora_scale, seed)
|
309 |
yield final_image, seed, gr.update(visible=False)
|
310 |
else:
|
311 |
+
# Text-to-Image Generation with Previews
|
312 |
+
pipe.to("cuda")
|
313 |
+
generator = torch.Generator(device="cuda").manual_seed(seed)
|
314 |
+
|
315 |
+
# Callback for generating previews
|
316 |
+
def callback_on_step_end(pipe, step_index, timestep, callback_kwargs):
|
317 |
+
latents = callback_kwargs["latents"]
|
318 |
+
# Use the fast taef1 decoder for previews
|
319 |
+
with torch.no_grad():
|
320 |
+
image = pipe.decode_latents(latents.to(dtype))[0]
|
321 |
+
progress_bar = f'<div class="progress-container"><div class="progress-bar" style="--current: {step_index + 1}; --total: {steps};"></div></div>'
|
322 |
+
yield {"image": image, "seed": seed, "progress": gr.update(value=progress_bar, visible=True)}
|
323 |
+
return callback_kwargs
|
324 |
+
|
325 |
+
# Generate image with step-by-step previews
|
326 |
+
with Timer("Generating image with previews"):
|
327 |
+
generation_output = pipe(
|
328 |
+
prompt=prompt_mash,
|
329 |
+
num_inference_steps=steps,
|
330 |
+
true_cfg_scale=cfg_scale,
|
331 |
+
width=width,
|
332 |
+
height=height,
|
333 |
+
generator=generator,
|
334 |
+
output_type="latent", # Get latents to decode with the good VAE later
|
335 |
+
callback_on_step_end=callback_on_step_end
|
336 |
+
)
|
337 |
+
|
338 |
+
# Decode the final image with the high-quality VAE
|
339 |
+
with Timer("Final decoding with good VAE"):
|
340 |
+
final_latents = generation_output.images
|
341 |
+
pipe.vae = good_vae # Temporarily swap to the good VAE
|
342 |
+
final_image = pipe.decode_latents(final_latents.to(dtype))[0]
|
343 |
+
pipe.vae = taef1 # Swap back to taef1 for the next run
|
344 |
+
|
345 |
+
yield final_image, seed, gr.update(visible=False)
|
346 |
+
|
347 |
|
348 |
def fetch_hf_adapter_files(link):
|
349 |
split_link = link.split("/")
|
|
|
352 |
|
353 |
print(f"Repository attempted: {split_link}")
|
354 |
|
|
|
355 |
model_card = ModelCard.load(link)
|
356 |
base_model = model_card.data.get("base_model")
|
357 |
print(f"Base model: {base_model}")
|
358 |
|
|
|
359 |
acceptable_models = {"Qwen/Qwen-Image"}
|
|
|
360 |
models_to_check = base_model if isinstance(base_model, list) else [base_model]
|
361 |
|
362 |
if not any(model in acceptable_models for model in models_to_check):
|
363 |
raise Exception("Not a Qwen-Image LoRA!")
|
364 |
|
365 |
+
image_path = model_card.data.get("widget", [{}])[0].get("output", {}).get("url")
|
|
|
366 |
trigger_word = model_card.data.get("instance_prompt", "")
|
367 |
image_url = f"https://huggingface.co/{link}/resolve/main/{image_path}" if image_path else None
|
368 |
|
|
|
369 |
fs = HfFileSystem()
|
370 |
try:
|
371 |
list_of_files = fs.ls(link, detail=False)
|
372 |
+
safetensors_name = next((f.split('/')[-1] for f in list_of_files if f.endswith(".safetensors")), None)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
373 |
if not safetensors_name:
|
374 |
raise Exception("No valid *.safetensors file found in the repository.")
|
|
|
375 |
except Exception as e:
|
376 |
print(e)
|
377 |
+
raise Exception("Could not find a valid *.safetensors file in the Hugging Face repository.")
|
378 |
|
379 |
return split_link[1], link, safetensors_name, trigger_word, image_url
|
380 |
|
381 |
def validate_custom_adapter(link):
|
382 |
print(f"Checking a custom model on: {link}")
|
383 |
+
if link.startswith("https://huggingface.co"):
|
384 |
+
link = urlparse(link).path.strip("/")
|
385 |
+
return fetch_hf_adapter_files(link)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
386 |
|
387 |
def incorporate_custom_adapter(custom_lora):
|
388 |
global loras
|
|
|
402 |
</div>
|
403 |
</div>
|
404 |
'''
|
405 |
+
existing_item_index = next((i for i, item in enumerate(loras) if item['repo'] == repo), None)
|
406 |
if existing_item_index is None:
|
407 |
+
new_item = {"image": image, "title": title, "repo": repo, "weights": path, "trigger_word": trigger_word}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
408 |
loras.append(new_item)
|
409 |
+
existing_item_index = len(loras) - 1
|
410 |
|
411 |
return gr.update(visible=True, value=card), gr.update(visible=True), gr.Gallery(selected_index=None), f"Custom: {path}", existing_item_index, trigger_word
|
412 |
except Exception as e:
|
413 |
+
gr.Warning(f"Invalid LoRA: {e}")
|
414 |
+
return gr.update(visible=True, value=f"Invalid LoRA: {e}"), gr.update(visible=True), gr.update(), "", None, ""
|
415 |
+
return gr.update(visible=False), gr.update(visible=False), gr.update(), "", None, ""
|
|
|
416 |
|
417 |
def discard_custom_adapter():
|
418 |
return gr.update(visible=False), gr.update(visible=False), gr.update(), "", None, ""
|
419 |
|
420 |
+
|
421 |
css = '''
|
422 |
#gen_btn{height: 100%}
|
423 |
#gen_column{align-self: stretch}
|
|
|
433 |
#progress{height:30px}
|
434 |
#progress .generating{display:none}
|
435 |
.progress-container {width: 100%;height: 30px;background-color: #f0f0f0;border-radius: 15px;overflow: hidden;margin-bottom: 20px}
|
436 |
+
.progress-bar {height: 100%;background-color: #4f46e5;width: calc(var(--current) / var(--total) * 100%);transition: width 0.1s ease-in-out}
|
437 |
'''
|
438 |
|
439 |
with gr.Blocks(theme="bethecloud/storj_theme", css=css, delete_cache=(120, 120)) as app:
|
|
|
457 |
elem_id="gallery",
|
458 |
show_share_button=False
|
459 |
)
|
460 |
+
with gr.Accordion("Image-to-Image (Optional)", open=False):
|
461 |
+
image_input = gr.Image(type="filepath", label="Input Image")
|
462 |
+
image_strength = gr.Slider(label="Image Strength", minimum=0.1, maximum=1.0, step=0.05, value=0.6)
|
463 |
+
|
464 |
with gr.Group():
|
465 |
custom_lora = gr.Textbox(label="Custom LoRA", info="LoRA Hugging Face path", placeholder="username/lora-model-name")
|
466 |
gr.Markdown("[Check Qwen-Image LoRAs](https://huggingface.co/models?other=base_model:adapter:Qwen/Qwen-Image)", elem_id="lora_list")
|
|
|
469 |
|
470 |
with gr.Column():
|
471 |
result = gr.Image(label="Generated Image")
|
472 |
+
progress_bar = gr.HTML(visible=False, elem_id="progress")
|
473 |
|
474 |
with gr.Row():
|
475 |
aspect_ratio = gr.Dropdown(
|
476 |
label="Aspect Ratio",
|
477 |
choices=["1:1", "16:9", "9:16", "4:3", "3:4", "3:2", "2:3"],
|
478 |
value="1:1"
|
479 |
+
)
|
480 |
with gr.Row():
|
481 |
speed_mode = gr.Dropdown(
|
482 |
label="Output Mode",
|
|
|
491 |
with gr.Column():
|
492 |
with gr.Row():
|
493 |
cfg_scale = gr.Slider(
|
494 |
+
label="Guidance Scale",
|
495 |
minimum=1.0,
|
496 |
maximum=5.0,
|
497 |
step=0.1,
|
498 |
value=4.0,
|
499 |
+
info="Lower for speed mode, higher for quality. Also called 'True CFG'."
|
500 |
)
|
501 |
steps = gr.Slider(
|
502 |
label="Steps",
|
|
|
511 |
randomize_seed = gr.Checkbox(True, label="Randomize seed")
|
512 |
seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0, randomize=True)
|
513 |
lora_scale = gr.Slider(label="LoRA Scale", minimum=0, maximum=2, step=0.01, value=1.0)
|
|
|
|
|
|
|
|
|
514 |
|
515 |
# Event handlers
|
516 |
gallery.select(
|
|
|
536 |
outputs=[custom_lora_info, custom_lora_button, gallery, selected_info, selected_index, custom_lora]
|
537 |
)
|
538 |
|
539 |
+
gen_inputs = [prompt, image_input, image_strength, cfg_scale, steps, selected_index, randomize_seed, seed, aspect_ratio, lora_scale, speed_mode]
|
540 |
+
gen_outputs = [result, seed, progress_bar]
|
541 |
+
|
542 |
+
generate_button.click(
|
543 |
+
fn=process_generation_request,
|
544 |
+
inputs=gen_inputs,
|
545 |
+
outputs=gen_outputs
|
546 |
+
)
|
547 |
+
prompt.submit(
|
548 |
+
fn=process_generation_request,
|
549 |
+
inputs=gen_inputs,
|
550 |
+
outputs=gen_outputs
|
551 |
)
|
552 |
|
553 |
app.queue()
|