Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -24,31 +24,6 @@ import shutil
|
|
24 |
import uuid
|
25 |
import zipfile
|
26 |
|
27 |
-
# Helper functions
|
28 |
-
def save_image(img):
|
29 |
-
unique_name = str(uuid.uuid4()) + ".png"
|
30 |
-
img.save(unique_name)
|
31 |
-
return unique_name
|
32 |
-
|
33 |
-
MAX_SEED = np.iinfo(np.int32).max
|
34 |
-
MAX_IMAGE_SIZE = 2048
|
35 |
-
|
36 |
-
# Load Qwen/Qwen-Image pipeline
|
37 |
-
dtype = torch.bfloat16
|
38 |
-
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
39 |
-
|
40 |
-
# Load Qwen model
|
41 |
-
pipe = DiffusionPipeline.from_pretrained("Qwen/Qwen-Image", torch_dtype=dtype).to(device)
|
42 |
-
|
43 |
-
# Aspect ratios
|
44 |
-
aspect_ratios = {
|
45 |
-
"1:1": (1328, 1328),
|
46 |
-
"16:9": (1664, 928),
|
47 |
-
"9:16": (928, 1664),
|
48 |
-
"4:3": (1472, 1140),
|
49 |
-
"3:4": (1140, 1472)
|
50 |
-
}
|
51 |
-
|
52 |
loras = [
|
53 |
# Sample Qwen-compatible LoRAs
|
54 |
{
|
@@ -88,84 +63,271 @@ loras = [
|
|
88 |
},
|
89 |
]
|
90 |
|
91 |
-
|
92 |
-
|
93 |
-
|
94 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
95 |
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
100 |
|
101 |
-
if
|
102 |
-
|
103 |
-
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
112 |
|
113 |
-
|
114 |
-
tmp_dir = tempfile.mkdtemp()
|
115 |
-
local_path = os.path.join(tmp_dir, os.path.basename(urlparse(url).path))
|
116 |
-
try:
|
117 |
-
print(f"Downloading LoRA from {url}...")
|
118 |
-
resp = requests.get(url, stream=True)
|
119 |
-
resp.raise_for_status()
|
120 |
-
with open(local_path, "wb") as f:
|
121 |
-
for chunk in resp.iter_content(chunk_size=8192):
|
122 |
-
f.write(chunk)
|
123 |
-
print(f"Saved LoRA to {local_path}")
|
124 |
-
pipe.load_lora_weights(local_path, adapter_name="default")
|
125 |
-
finally:
|
126 |
-
shutil.rmtree(tmp_dir, ignore_errors=True)
|
127 |
|
128 |
-
|
129 |
-
|
130 |
-
if
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
135 |
|
136 |
-
#
|
137 |
-
|
138 |
-
|
139 |
-
|
140 |
-
|
141 |
-
|
142 |
-
|
143 |
-
|
144 |
-
image_url = model_info["cardData"]["widget"][0].get("output", {}).get("url", None)
|
145 |
|
146 |
-
#
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
152 |
-
|
153 |
-
|
154 |
-
|
155 |
-
|
|
|
156 |
|
157 |
-
|
158 |
-
|
159 |
-
|
160 |
-
|
161 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
162 |
|
163 |
def check_custom_model(link):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
164 |
if link.startswith("https://"):
|
165 |
if link.startswith("https://huggingface.co") or link.startswith("https://www.huggingface.co"):
|
166 |
link_split = link.split("huggingface.co/")
|
167 |
return get_huggingface_safetensors(link_split[1])
|
168 |
-
else:
|
169 |
return get_huggingface_safetensors(link)
|
170 |
|
171 |
def add_custom_lora(custom_lora):
|
@@ -173,9 +335,6 @@ def add_custom_lora(custom_lora):
|
|
173 |
if custom_lora:
|
174 |
try:
|
175 |
title, repo, path, trigger_word, image = check_custom_model(custom_lora)
|
176 |
-
if not title:
|
177 |
-
raise Exception("Invalid LoRA model")
|
178 |
-
|
179 |
print(f"Loaded custom LoRA: {repo}")
|
180 |
card = f'''
|
181 |
<div class="custom_lora_card">
|
@@ -189,9 +348,8 @@ def add_custom_lora(custom_lora):
|
|
189 |
</div>
|
190 |
</div>
|
191 |
'''
|
192 |
-
|
193 |
existing_item_index = next((index for (index, item) in enumerate(loras) if item['repo'] == repo), None)
|
194 |
-
if
|
195 |
new_item = {
|
196 |
"image": image,
|
197 |
"title": title,
|
@@ -199,161 +357,21 @@ def add_custom_lora(custom_lora):
|
|
199 |
"weights": path,
|
200 |
"trigger_word": trigger_word
|
201 |
}
|
202 |
-
|
203 |
loras.append(new_item)
|
|
|
204 |
|
205 |
return gr.update(visible=True, value=card), gr.update(visible=True), gr.Gallery(selected_index=None), f"Custom: {path}", existing_item_index, trigger_word
|
206 |
except Exception as e:
|
207 |
-
gr.Warning(f"Invalid LoRA: either you entered an invalid link, or a non-Qwen LoRA")
|
208 |
-
return gr.update(visible=True, value=f"Invalid LoRA: either you entered an invalid link, a non-Qwen LoRA"), gr.update(visible=
|
209 |
else:
|
210 |
return gr.update(visible=False), gr.update(visible=False), gr.update(), "", None, ""
|
211 |
|
212 |
def remove_custom_lora():
|
213 |
return gr.update(visible=False), gr.update(visible=False), gr.update(), "", None, ""
|
214 |
|
215 |
-
|
216 |
-
selected_lora = loras[evt.index]
|
217 |
-
new_placeholder = f"Type a prompt for {selected_lora['title']}"
|
218 |
-
lora_repo = selected_lora["repo"]
|
219 |
-
updated_text = f"### Selected: [{lora_repo}](https://huggingface.co/{lora_repo}) ✅"
|
220 |
-
|
221 |
-
# Update aspect ratio based on LoRA if it has aspect info
|
222 |
-
if "aspect" in selected_lora:
|
223 |
-
if selected_lora["aspect"] == "portrait":
|
224 |
-
width = 928
|
225 |
-
height = 1664
|
226 |
-
elif selected_lora["aspect"] == "landscape":
|
227 |
-
width = 1664
|
228 |
-
height = 928
|
229 |
-
else:
|
230 |
-
width = 1328
|
231 |
-
height = 1328
|
232 |
-
|
233 |
-
return (
|
234 |
-
gr.update(placeholder=new_placeholder),
|
235 |
-
updated_text,
|
236 |
-
evt.index,
|
237 |
-
width,
|
238 |
-
height,
|
239 |
-
)
|
240 |
-
|
241 |
-
@spaces.GPU(duration=120)
|
242 |
-
def generate_qwen(
|
243 |
-
prompt: str,
|
244 |
-
negative_prompt: str = "",
|
245 |
-
seed: int = 0,
|
246 |
-
width: int = 1024,
|
247 |
-
height: int = 1024,
|
248 |
-
guidance_scale: float = 4.0,
|
249 |
-
randomize_seed: bool = False,
|
250 |
-
num_inference_steps: int = 50,
|
251 |
-
num_images: int = 1,
|
252 |
-
zip_images: bool = False,
|
253 |
-
lora_input: str = "",
|
254 |
-
lora_scale: float = 1.0,
|
255 |
-
progress=gr.Progress(track_tqdm=True),
|
256 |
-
):
|
257 |
-
if randomize_seed:
|
258 |
-
seed = random.randint(0, MAX_SEED)
|
259 |
-
|
260 |
-
generator = torch.Generator(device).manual_seed(seed)
|
261 |
-
|
262 |
-
start_time = time.time()
|
263 |
-
|
264 |
-
# Clear any existing LoRA adapters
|
265 |
-
current_adapters = pipe.get_list_adapters()
|
266 |
-
for adapter in current_adapters:
|
267 |
-
pipe.delete_adapters(adapter)
|
268 |
-
pipe.disable_lora()
|
269 |
-
|
270 |
-
use_lora = False
|
271 |
-
if lora_input and lora_input.strip() != "":
|
272 |
-
load_lora_opt(pipe, lora_input)
|
273 |
-
pipe.set_adapters(["default"], adapter_weights=[lora_scale])
|
274 |
-
use_lora = True
|
275 |
-
|
276 |
-
images = pipe(
|
277 |
-
prompt=prompt,
|
278 |
-
negative_prompt=negative_prompt if negative_prompt else "",
|
279 |
-
height=height,
|
280 |
-
width=width,
|
281 |
-
guidance_scale=guidance_scale,
|
282 |
-
num_inference_steps=num_inference_steps,
|
283 |
-
num_images_per_prompt=num_images,
|
284 |
-
generator=generator,
|
285 |
-
output_type="pil",
|
286 |
-
).images
|
287 |
-
|
288 |
-
end_time = time.time()
|
289 |
-
duration = end_time - start_time
|
290 |
-
|
291 |
-
image_paths = [save_image(img) for img in images]
|
292 |
-
zip_path = None
|
293 |
-
if zip_images:
|
294 |
-
zip_name = str(uuid.uuid4()) + ".zip"
|
295 |
-
with zipfile.ZipFile(zip_name, 'w') as zipf:
|
296 |
-
for i, img_path in enumerate(image_paths):
|
297 |
-
zipf.write(img_path, arcname=f"Img_{i}.png")
|
298 |
-
zip_path = zip_name
|
299 |
-
|
300 |
-
# Clean up adapters
|
301 |
-
current_adapters = pipe.get_list_adapters()
|
302 |
-
for adapter in current_adapters:
|
303 |
-
pipe.delete_adapters(adapter)
|
304 |
-
pipe.disable_lora()
|
305 |
-
|
306 |
-
return image_paths, seed, f"{duration:.2f}", zip_path
|
307 |
-
|
308 |
-
@spaces.GPU(duration=120)
|
309 |
-
def run_lora(
|
310 |
-
prompt: str,
|
311 |
-
negative_prompt: str,
|
312 |
-
use_negative_prompt: bool,
|
313 |
-
seed: int,
|
314 |
-
width: int,
|
315 |
-
height: int,
|
316 |
-
guidance_scale: float,
|
317 |
-
randomize_seed: bool,
|
318 |
-
num_inference_steps: int,
|
319 |
-
num_images: int,
|
320 |
-
zip_images: bool,
|
321 |
-
selected_index: int,
|
322 |
-
lora_scale: float,
|
323 |
-
progress=gr.Progress(track_tqdm=True),
|
324 |
-
):
|
325 |
-
if selected_index is None:
|
326 |
-
raise gr.Error("You must select a LoRA before proceeding.🧨")
|
327 |
-
|
328 |
-
selected_lora = loras[selected_index]
|
329 |
-
lora_repo = selected_lora["repo"]
|
330 |
-
trigger_word = selected_lora["trigger_word"]
|
331 |
-
|
332 |
-
if trigger_word:
|
333 |
-
prompt_mash = f"{trigger_word} {prompt}"
|
334 |
-
else:
|
335 |
-
prompt_mash = prompt
|
336 |
-
|
337 |
-
final_negative_prompt = negative_prompt if use_negative_prompt else ""
|
338 |
-
|
339 |
-
if randomize_seed:
|
340 |
-
seed = random.randint(0, MAX_SEED)
|
341 |
-
|
342 |
-
return generate_qwen(
|
343 |
-
prompt=prompt_mash,
|
344 |
-
negative_prompt=final_negative_prompt,
|
345 |
-
seed=seed,
|
346 |
-
width=width,
|
347 |
-
height=height,
|
348 |
-
guidance_scale=guidance_scale,
|
349 |
-
randomize_seed=False, # Already handled
|
350 |
-
num_inference_steps=num_inference_steps,
|
351 |
-
num_images=num_images,
|
352 |
-
zip_images=zip_images,
|
353 |
-
lora_input=lora_repo,
|
354 |
-
lora_scale=lora_scale,
|
355 |
-
progress=progress,
|
356 |
-
)
|
357 |
|
358 |
css = '''
|
359 |
#gen_btn{height: 100%}
|
@@ -366,10 +384,7 @@ css = '''
|
|
366 |
.card_internal{display: flex;height: 100px;margin-top: .5em}
|
367 |
.card_internal img{margin-right: 1em}
|
368 |
.styler{--form-gap-width: 0px !important}
|
369 |
-
#
|
370 |
-
#progress .generating{display:none}
|
371 |
-
.progress-container {width: 100%;height: 30px;background-color: #f0f0f0;border-radius: 15px;overflow: hidden;margin-bottom: 20px}
|
372 |
-
.progress-bar {height: 100%;background-color: #4f46e5;width: calc(var(--current) / var(--total) * 100%);transition: width 0.5s ease-in-out}
|
373 |
'''
|
374 |
|
375 |
with gr.Blocks(theme="bethecloud/storj_theme", css=css, delete_cache=(120, 120)) as app:
|
@@ -378,7 +393,7 @@ with gr.Blocks(theme="bethecloud/storj_theme", css=css, delete_cache=(120, 120))
|
|
378 |
|
379 |
with gr.Row():
|
380 |
with gr.Column(scale=3):
|
381 |
-
prompt = gr.Textbox(label="Prompt", lines=1, placeholder="
|
382 |
with gr.Column(scale=1, elem_id="gen_column"):
|
383 |
generate_button = gr.Button("Generate", variant="primary", elem_id="gen_btn")
|
384 |
|
@@ -387,89 +402,73 @@ with gr.Blocks(theme="bethecloud/storj_theme", css=css, delete_cache=(120, 120))
|
|
387 |
selected_info = gr.Markdown("")
|
388 |
gallery = gr.Gallery(
|
389 |
[(item["image"], item["title"]) for item in loras],
|
390 |
-
label="
|
391 |
allow_preview=False,
|
392 |
columns=3,
|
393 |
elem_id="gallery",
|
394 |
show_share_button=False
|
395 |
)
|
396 |
with gr.Group():
|
397 |
-
custom_lora = gr.Textbox(label="
|
398 |
-
gr.Markdown("[Check
|
399 |
custom_lora_info = gr.HTML(visible=False)
|
400 |
custom_lora_button = gr.Button("Remove custom LoRA", visible=False)
|
401 |
|
402 |
with gr.Column():
|
403 |
-
result = gr.
|
|
|
404 |
with gr.Row():
|
405 |
aspect_ratio = gr.Dropdown(
|
406 |
label="Aspect Ratio",
|
407 |
-
choices=
|
408 |
-
value="1:1"
|
409 |
-
|
410 |
with gr.Row():
|
411 |
-
|
412 |
-
|
413 |
-
|
414 |
-
|
415 |
-
|
416 |
-
with gr.Row():
|
417 |
-
use_negative_prompt = gr.Checkbox(
|
418 |
-
label="Use negative prompt",
|
419 |
-
value=True,
|
420 |
)
|
421 |
-
negative_prompt = gr.Text(
|
422 |
-
label="Negative prompt",
|
423 |
-
max_lines=1,
|
424 |
-
placeholder="Enter a negative prompt",
|
425 |
-
value="text, watermark, copyright, blurry, low resolution",
|
426 |
-
)
|
427 |
-
|
428 |
-
with gr.Row():
|
429 |
-
cfg_scale = gr.Slider(label="CFG Scale", minimum=1, maximum=20, step=0.5, value=4.0)
|
430 |
-
steps = gr.Slider(label="Steps", minimum=1, maximum=100, step=1, value=50)
|
431 |
-
|
432 |
-
with gr.Row():
|
433 |
-
width = gr.Slider(label="Width", minimum=512, maximum=2048, step=64, value=1328)
|
434 |
-
height = gr.Slider(label="Height", minimum=512, maximum=2048, step=64, value=1328)
|
435 |
-
|
436 |
-
with gr.Row():
|
437 |
-
num_images = gr.Slider(label="Number of Images", minimum=1, maximum=5, step=1, value=1)
|
438 |
-
zip_images = gr.Checkbox(label="Zip generated images", value=False)
|
439 |
|
440 |
-
|
441 |
-
|
442 |
-
seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0, randomize=True)
|
443 |
-
lora_scale = gr.Slider(label="LoRA Scale", minimum=0, maximum=2, step=0.01, value=1.0)
|
444 |
-
|
445 |
-
# Output information
|
446 |
with gr.Row():
|
447 |
-
|
448 |
-
|
449 |
-
|
450 |
-
|
451 |
-
|
452 |
-
|
453 |
-
|
454 |
-
|
455 |
-
|
456 |
-
|
457 |
-
|
458 |
-
|
459 |
-
|
460 |
-
|
461 |
-
|
462 |
-
|
463 |
-
|
464 |
-
|
465 |
-
|
466 |
-
|
467 |
-
|
468 |
-
|
|
|
|
|
|
|
|
|
469 |
gallery.select(
|
470 |
update_selection,
|
471 |
-
inputs=[
|
472 |
-
outputs=[prompt, selected_info, selected_index,
|
|
|
|
|
|
|
|
|
|
|
|
|
473 |
)
|
474 |
|
475 |
custom_lora.input(
|
@@ -486,22 +485,8 @@ with gr.Blocks(theme="bethecloud/storj_theme", css=css, delete_cache=(120, 120))
|
|
486 |
gr.on(
|
487 |
triggers=[generate_button.click, prompt.submit],
|
488 |
fn=run_lora,
|
489 |
-
inputs=[
|
490 |
-
|
491 |
-
negative_prompt,
|
492 |
-
use_negative_prompt,
|
493 |
-
seed,
|
494 |
-
width,
|
495 |
-
height,
|
496 |
-
#guidance_scale,
|
497 |
-
randomize_seed,
|
498 |
-
steps,
|
499 |
-
num_images,
|
500 |
-
zip_images,
|
501 |
-
selected_index,
|
502 |
-
lora_scale,
|
503 |
-
],
|
504 |
-
outputs=[result, seed_display, generation_time, zip_file]
|
505 |
)
|
506 |
|
507 |
app.queue()
|
|
|
24 |
import uuid
|
25 |
import zipfile
|
26 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
27 |
loras = [
|
28 |
# Sample Qwen-compatible LoRAs
|
29 |
{
|
|
|
63 |
},
|
64 |
]
|
65 |
|
66 |
+
# Initialize the base model
|
67 |
+
dtype = torch.bfloat16
|
68 |
+
device = "cuda" if torch.cuda.is_available() else "cpu"
|
69 |
+
base_model = "Qwen/Qwen-Image"
|
70 |
+
|
71 |
+
# Scheduler configuration from the Qwen-Image-Lightning repository
|
72 |
+
scheduler_config = {
|
73 |
+
"base_image_seq_len": 256,
|
74 |
+
"base_shift": math.log(3),
|
75 |
+
"invert_sigmas": False,
|
76 |
+
"max_image_seq_len": 8192,
|
77 |
+
"max_shift": math.log(3),
|
78 |
+
"num_train_timesteps": 1000,
|
79 |
+
"shift": 1.0,
|
80 |
+
"shift_terminal": None,
|
81 |
+
"stochastic_sampling": False,
|
82 |
+
"time_shift_type": "exponential",
|
83 |
+
"use_beta_sigmas": False,
|
84 |
+
"use_dynamic_shifting": True,
|
85 |
+
"use_exponential_sigmas": False,
|
86 |
+
"use_karras_sigmas": False,
|
87 |
+
}
|
88 |
+
|
89 |
+
scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config)
|
90 |
+
pipe = DiffusionPipeline.from_pretrained(
|
91 |
+
base_model, scheduler=scheduler, torch_dtype=dtype
|
92 |
+
).to(device)
|
93 |
+
|
94 |
+
# Lightning LoRA info (no global state)
|
95 |
+
LIGHTNING_LORA_REPO = "lightx2v/Qwen-Image-Lightning"
|
96 |
+
LIGHTNING_LORA_WEIGHT = "Qwen-Image-Lightning-8steps-V1.0.safetensors"
|
97 |
+
|
98 |
+
MAX_SEED = np.iinfo(np.int32).max
|
99 |
+
|
100 |
+
class calculateDuration:
|
101 |
+
def __init__(self, activity_name=""):
|
102 |
+
self.activity_name = activity_name
|
103 |
+
|
104 |
+
def __enter__(self):
|
105 |
+
self.start_time = time.time()
|
106 |
+
return self
|
107 |
|
108 |
+
def __exit__(self, exc_type, exc_value, traceback):
|
109 |
+
self.end_time = time.time()
|
110 |
+
self.elapsed_time = self.end_time - self.start_time
|
111 |
+
if self.activity_name:
|
112 |
+
print(f"Elapsed time for {self.activity_name}: {self.elapsed_time:.6f} seconds")
|
113 |
+
else:
|
114 |
+
print(f"Elapsed time: {self.elapsed_time:.6f} seconds")
|
115 |
+
|
116 |
+
def get_image_size(aspect_ratio):
|
117 |
+
"""Converts aspect ratio string to width, height tuple."""
|
118 |
+
if aspect_ratio == "1:1":
|
119 |
+
return 1024, 1024
|
120 |
+
elif aspect_ratio == "16:9":
|
121 |
+
return 1152, 640
|
122 |
+
elif aspect_ratio == "9:16":
|
123 |
+
return 640, 1152
|
124 |
+
elif aspect_ratio == "4:3":
|
125 |
+
return 1024, 768
|
126 |
+
elif aspect_ratio == "3:4":
|
127 |
+
return 768, 1024
|
128 |
+
elif aspect_ratio == "3:2":
|
129 |
+
return 1024, 688
|
130 |
+
elif aspect_ratio == "2:3":
|
131 |
+
return 688, 1024
|
132 |
+
else:
|
133 |
+
return 1024, 1024
|
134 |
+
|
135 |
+
def update_selection(evt: gr.SelectData, aspect_ratio):
|
136 |
+
selected_lora = loras[evt.index]
|
137 |
+
new_placeholder = f"Type a prompt for {selected_lora['title']}"
|
138 |
+
lora_repo = selected_lora["repo"]
|
139 |
+
updated_text = f"### Selected: [{lora_repo}](https://huggingface.co/{lora_repo}) ✨"
|
140 |
|
141 |
+
# Update aspect ratio if specified in LoRA config
|
142 |
+
if "aspect" in selected_lora:
|
143 |
+
if selected_lora["aspect"] == "portrait":
|
144 |
+
aspect_ratio = "9:16"
|
145 |
+
elif selected_lora["aspect"] == "landscape":
|
146 |
+
aspect_ratio = "16:9"
|
147 |
+
else:
|
148 |
+
aspect_ratio = "1:1"
|
149 |
+
|
150 |
+
return (
|
151 |
+
gr.update(placeholder=new_placeholder),
|
152 |
+
updated_text,
|
153 |
+
evt.index,
|
154 |
+
aspect_ratio,
|
155 |
+
)
|
156 |
+
|
157 |
+
def handle_speed_mode(speed_mode):
|
158 |
+
"""Update UI based on speed/quality toggle."""
|
159 |
+
if speed_mode == "Speed (8 steps)":
|
160 |
+
return gr.update(value="Speed mode selected - 8 steps with Lightning LoRA"), 8, 1.0
|
161 |
+
else:
|
162 |
+
return gr.update(value="Quality mode selected - 45 steps for best quality"), 45, 3.5
|
163 |
+
|
164 |
+
@spaces.GPU(duration=70)
|
165 |
+
def generate_image(prompt_mash, steps, seed, cfg_scale, width, height, lora_scale, negative_prompt=""):
|
166 |
+
pipe.to("cuda")
|
167 |
+
generator = torch.Generator(device="cuda").manual_seed(seed)
|
168 |
+
|
169 |
+
with calculateDuration("Generating image"):
|
170 |
+
# Generate image
|
171 |
+
image = pipe(
|
172 |
+
prompt=prompt_mash,
|
173 |
+
negative_prompt=negative_prompt,
|
174 |
+
num_inference_steps=steps,
|
175 |
+
true_cfg_scale=cfg_scale, # Use true_cfg_scale for Qwen-Image
|
176 |
+
width=width,
|
177 |
+
height=height,
|
178 |
+
generator=generator,
|
179 |
+
).images[0]
|
180 |
|
181 |
+
return image
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
182 |
|
183 |
+
@spaces.GPU(duration=70)
|
184 |
+
def run_lora(prompt, cfg_scale, steps, selected_index, randomize_seed, seed, aspect_ratio, lora_scale, speed_mode, progress=gr.Progress(track_tqdm=True)):
|
185 |
+
if selected_index is None:
|
186 |
+
raise gr.Error("You must select a LoRA before proceeding.")
|
187 |
+
|
188 |
+
selected_lora = loras[selected_index]
|
189 |
+
lora_path = selected_lora["repo"]
|
190 |
+
trigger_word = selected_lora["trigger_word"]
|
191 |
+
|
192 |
+
# Prepare prompt with trigger word
|
193 |
+
if trigger_word:
|
194 |
+
if "trigger_position" in selected_lora:
|
195 |
+
if selected_lora["trigger_position"] == "prepend":
|
196 |
+
prompt_mash = f"{trigger_word} {prompt}"
|
197 |
+
else:
|
198 |
+
prompt_mash = f"{prompt} {trigger_word}"
|
199 |
+
else:
|
200 |
+
prompt_mash = f"{trigger_word} {prompt}"
|
201 |
+
else:
|
202 |
+
prompt_mash = prompt
|
203 |
+
|
204 |
+
# Always unload any existing LoRAs first to avoid conflicts
|
205 |
+
with calculateDuration("Unloading existing LoRAs"):
|
206 |
+
pipe.unload_lora_weights()
|
207 |
+
|
208 |
+
# Load LoRAs based on speed mode
|
209 |
+
if speed_mode == "Speed (8 steps)":
|
210 |
+
with calculateDuration("Loading Lightning LoRA and style LoRA"):
|
211 |
+
# Load Lightning LoRA first
|
212 |
+
pipe.load_lora_weights(
|
213 |
+
LIGHTNING_LORA_REPO,
|
214 |
+
weight_name=LIGHTNING_LORA_WEIGHT,
|
215 |
+
adapter_name="lightning"
|
216 |
+
)
|
217 |
|
218 |
+
# Load the selected style LoRA
|
219 |
+
weight_name = selected_lora.get("weights", None)
|
220 |
+
pipe.load_lora_weights(
|
221 |
+
lora_path,
|
222 |
+
weight_name=weight_name,
|
223 |
+
low_cpu_mem_usage=True,
|
224 |
+
adapter_name="style"
|
225 |
+
)
|
|
|
226 |
|
227 |
+
# Set both adapters active with their weights
|
228 |
+
pipe.set_adapters(["lightning", "style"], adapter_weights=[1.0, lora_scale])
|
229 |
+
else:
|
230 |
+
# Quality mode - only load the style LoRA
|
231 |
+
with calculateDuration(f"Loading LoRA weights for {selected_lora['title']}"):
|
232 |
+
weight_name = selected_lora.get("weights", None)
|
233 |
+
pipe.load_lora_weights(
|
234 |
+
lora_path,
|
235 |
+
weight_name=weight_name,
|
236 |
+
low_cpu_mem_usage=True
|
237 |
+
)
|
238 |
|
239 |
+
# Set random seed for reproducibility
|
240 |
+
with calculateDuration("Randomizing seed"):
|
241 |
+
if randomize_seed:
|
242 |
+
seed = random.randint(0, MAX_SEED)
|
243 |
+
|
244 |
+
# Get image dimensions from aspect ratio
|
245 |
+
width, height = get_image_size(aspect_ratio)
|
246 |
+
|
247 |
+
# Generate the image
|
248 |
+
final_image = generate_image(prompt_mash, steps, seed, cfg_scale, width, height, lora_scale)
|
249 |
+
|
250 |
+
return final_image, seed
|
251 |
+
|
252 |
+
def get_huggingface_safetensors(link):
|
253 |
+
split_link = link.split("/")
|
254 |
+
if len(split_link) != 2:
|
255 |
+
raise Exception("Invalid Hugging Face repository link format.")
|
256 |
+
|
257 |
+
print(f"Repository attempted: {split_link}")
|
258 |
+
|
259 |
+
# Load model card
|
260 |
+
model_card = ModelCard.load(link)
|
261 |
+
base_model = model_card.data.get("base_model")
|
262 |
+
print(f"Base model: {base_model}")
|
263 |
+
|
264 |
+
# Validate model type (for Qwen-Image)
|
265 |
+
acceptable_models = {"Qwen/Qwen-Image"}
|
266 |
+
|
267 |
+
models_to_check = base_model if isinstance(base_model, list) else [base_model]
|
268 |
+
|
269 |
+
if not any(model in acceptable_models for model in models_to_check):
|
270 |
+
raise Exception("Not a Qwen-Image LoRA!")
|
271 |
+
|
272 |
+
# Extract image and trigger word
|
273 |
+
image_path = model_card.data.get("widget", [{}])[0].get("output", {}).get("url", None)
|
274 |
+
trigger_word = model_card.data.get("instance_prompt", "")
|
275 |
+
image_url = f"https://huggingface.co/{link}/resolve/main/{image_path}" if image_path else None
|
276 |
+
|
277 |
+
# Initialize Hugging Face file system
|
278 |
+
fs = HfFileSystem()
|
279 |
+
try:
|
280 |
+
list_of_files = fs.ls(link, detail=False)
|
281 |
+
|
282 |
+
# Find safetensors file
|
283 |
+
safetensors_name = None
|
284 |
+
for file in list_of_files:
|
285 |
+
filename = file.split("/")[-1]
|
286 |
+
if filename.endswith(".safetensors"):
|
287 |
+
safetensors_name = filename
|
288 |
+
break
|
289 |
+
|
290 |
+
if not safetensors_name:
|
291 |
+
raise Exception("No valid *.safetensors file found in the repository.")
|
292 |
+
|
293 |
+
except Exception as e:
|
294 |
+
print(e)
|
295 |
+
raise Exception("You didn't include a valid Hugging Face repository with a *.safetensors LoRA")
|
296 |
+
|
297 |
+
return split_link[1], link, safetensors_name, trigger_word, image_url
|
298 |
|
299 |
def check_custom_model(link):
|
300 |
+
print(f"Checking a custom model on: {link}")
|
301 |
+
|
302 |
+
if link.endswith('.safetensors'):
|
303 |
+
if 'huggingface.co' in link:
|
304 |
+
parts = link.split('/')
|
305 |
+
try:
|
306 |
+
hf_index = parts.index('huggingface.co')
|
307 |
+
username = parts[hf_index + 1]
|
308 |
+
repo_name = parts[hf_index + 2]
|
309 |
+
repo = f"{username}/{repo_name}"
|
310 |
+
|
311 |
+
safetensors_name = parts[-1]
|
312 |
+
|
313 |
+
try:
|
314 |
+
model_card = ModelCard.load(repo)
|
315 |
+
trigger_word = model_card.data.get("instance_prompt", "")
|
316 |
+
image_path = model_card.data.get("widget", [{}])[0].get("output", {}).get("url", None)
|
317 |
+
image_url = f"https://huggingface.co/{repo}/resolve/main/{image_path}" if image_path else None
|
318 |
+
except:
|
319 |
+
trigger_word = ""
|
320 |
+
image_url = None
|
321 |
+
|
322 |
+
return repo_name, repo, safetensors_name, trigger_word, image_url
|
323 |
+
except:
|
324 |
+
raise Exception("Invalid safetensors URL format")
|
325 |
+
|
326 |
if link.startswith("https://"):
|
327 |
if link.startswith("https://huggingface.co") or link.startswith("https://www.huggingface.co"):
|
328 |
link_split = link.split("huggingface.co/")
|
329 |
return get_huggingface_safetensors(link_split[1])
|
330 |
+
else:
|
331 |
return get_huggingface_safetensors(link)
|
332 |
|
333 |
def add_custom_lora(custom_lora):
|
|
|
335 |
if custom_lora:
|
336 |
try:
|
337 |
title, repo, path, trigger_word, image = check_custom_model(custom_lora)
|
|
|
|
|
|
|
338 |
print(f"Loaded custom LoRA: {repo}")
|
339 |
card = f'''
|
340 |
<div class="custom_lora_card">
|
|
|
348 |
</div>
|
349 |
</div>
|
350 |
'''
|
|
|
351 |
existing_item_index = next((index for (index, item) in enumerate(loras) if item['repo'] == repo), None)
|
352 |
+
if existing_item_index is None:
|
353 |
new_item = {
|
354 |
"image": image,
|
355 |
"title": title,
|
|
|
357 |
"weights": path,
|
358 |
"trigger_word": trigger_word
|
359 |
}
|
360 |
+
print(new_item)
|
361 |
loras.append(new_item)
|
362 |
+
existing_item_index = len(loras) - 1 # Get the actual index after adding
|
363 |
|
364 |
return gr.update(visible=True, value=card), gr.update(visible=True), gr.Gallery(selected_index=None), f"Custom: {path}", existing_item_index, trigger_word
|
365 |
except Exception as e:
|
366 |
+
gr.Warning(f"Invalid LoRA: either you entered an invalid link, or a non-Qwen-Image LoRA, this was the issue: {e}")
|
367 |
+
return gr.update(visible=True, value=f"Invalid LoRA: either you entered an invalid link, a non-Qwen-Image LoRA"), gr.update(visible=True), gr.update(), "", None, ""
|
368 |
else:
|
369 |
return gr.update(visible=False), gr.update(visible=False), gr.update(), "", None, ""
|
370 |
|
371 |
def remove_custom_lora():
|
372 |
return gr.update(visible=False), gr.update(visible=False), gr.update(), "", None, ""
|
373 |
|
374 |
+
run_lora.zerogpu = True
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
375 |
|
376 |
css = '''
|
377 |
#gen_btn{height: 100%}
|
|
|
384 |
.card_internal{display: flex;height: 100px;margin-top: .5em}
|
385 |
.card_internal img{margin-right: 1em}
|
386 |
.styler{--form-gap-width: 0px !important}
|
387 |
+
#speed_status{padding: .5em; border-radius: 5px; margin: 1em 0}
|
|
|
|
|
|
|
388 |
'''
|
389 |
|
390 |
with gr.Blocks(theme="bethecloud/storj_theme", css=css, delete_cache=(120, 120)) as app:
|
|
|
393 |
|
394 |
with gr.Row():
|
395 |
with gr.Column(scale=3):
|
396 |
+
prompt = gr.Textbox(label="Prompt", lines=1, placeholder="Type a prompt after selecting a LoRA")
|
397 |
with gr.Column(scale=1, elem_id="gen_column"):
|
398 |
generate_button = gr.Button("Generate", variant="primary", elem_id="gen_btn")
|
399 |
|
|
|
402 |
selected_info = gr.Markdown("")
|
403 |
gallery = gr.Gallery(
|
404 |
[(item["image"], item["title"]) for item in loras],
|
405 |
+
label="LoRA Gallery",
|
406 |
allow_preview=False,
|
407 |
columns=3,
|
408 |
elem_id="gallery",
|
409 |
show_share_button=False
|
410 |
)
|
411 |
with gr.Group():
|
412 |
+
custom_lora = gr.Textbox(label="Custom LoRA", info="LoRA Hugging Face path", placeholder="username/qwen-image-custom-lora")
|
413 |
+
gr.Markdown("[Check Qwen-Image LoRAs](https://huggingface.co/models?other=base_model:adapter:Qwen/Qwen-Image)", elem_id="lora_list")
|
414 |
custom_lora_info = gr.HTML(visible=False)
|
415 |
custom_lora_button = gr.Button("Remove custom LoRA", visible=False)
|
416 |
|
417 |
with gr.Column():
|
418 |
+
result = gr.Image(label="Generated Image")
|
419 |
+
|
420 |
with gr.Row():
|
421 |
aspect_ratio = gr.Dropdown(
|
422 |
label="Aspect Ratio",
|
423 |
+
choices=["1:1", "16:9", "9:16", "4:3", "3:4", "3:2", "2:3"],
|
424 |
+
value="1:1"
|
425 |
+
)
|
426 |
with gr.Row():
|
427 |
+
speed_mode = gr.Dropdown(
|
428 |
+
label="Generation Mode",
|
429 |
+
choices=["Speed (8 steps)", "Quality (45 steps)"],
|
430 |
+
value="Quality (48 steps)",
|
|
|
|
|
|
|
|
|
|
|
431 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
432 |
|
433 |
+
speed_status = gr.Markdown("Quality mode active", elem_id="speed_status")
|
434 |
+
|
|
|
|
|
|
|
|
|
435 |
with gr.Row():
|
436 |
+
with gr.Accordion("Advanced Settings", open=False):
|
437 |
+
with gr.Column():
|
438 |
+
with gr.Row():
|
439 |
+
cfg_scale = gr.Slider(
|
440 |
+
label="Guidance Scale (True CFG)",
|
441 |
+
minimum=1.0,
|
442 |
+
maximum=5.0,
|
443 |
+
step=0.1,
|
444 |
+
value=3.5,
|
445 |
+
info="Lower for speed mode, higher for quality"
|
446 |
+
)
|
447 |
+
steps = gr.Slider(
|
448 |
+
label="Steps",
|
449 |
+
minimum=4,
|
450 |
+
maximum=50,
|
451 |
+
step=1,
|
452 |
+
value=45,
|
453 |
+
info="Automatically set by speed mode"
|
454 |
+
)
|
455 |
+
|
456 |
+
with gr.Row():
|
457 |
+
randomize_seed = gr.Checkbox(True, label="Randomize seed")
|
458 |
+
seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0, randomize=True)
|
459 |
+
lora_scale = gr.Slider(label="LoRA Scale", minimum=0, maximum=2, step=0.01, value=1.0)
|
460 |
+
|
461 |
+
# Event handlers
|
462 |
gallery.select(
|
463 |
update_selection,
|
464 |
+
inputs=[aspect_ratio],
|
465 |
+
outputs=[prompt, selected_info, selected_index, aspect_ratio]
|
466 |
+
)
|
467 |
+
|
468 |
+
speed_mode.change(
|
469 |
+
handle_speed_mode,
|
470 |
+
inputs=[speed_mode],
|
471 |
+
outputs=[speed_status, steps, cfg_scale]
|
472 |
)
|
473 |
|
474 |
custom_lora.input(
|
|
|
485 |
gr.on(
|
486 |
triggers=[generate_button.click, prompt.submit],
|
487 |
fn=run_lora,
|
488 |
+
inputs=[prompt, cfg_scale, steps, selected_index, randomize_seed, seed, aspect_ratio, lora_scale, speed_mode],
|
489 |
+
outputs=[result, seed]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
490 |
)
|
491 |
|
492 |
app.queue()
|