Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
@@ -13,7 +13,9 @@ import gradio as gr
|
|
13 |
import spaces
|
14 |
from diffusers import (
|
15 |
DiffusionPipeline,
|
16 |
-
FlowMatchEulerDiscreteScheduler
|
|
|
|
|
17 |
from huggingface_hub import (
|
18 |
hf_hub_download,
|
19 |
HfFileSystem,
|
@@ -142,15 +144,30 @@ pipe = DiffusionPipeline.from_pretrained(
|
|
142 |
base_model, scheduler=scheduler, torch_dtype=dtype
|
143 |
).to(device)
|
144 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
145 |
# Lightning LoRA info (no global state)
|
146 |
LIGHTNING_LORA_REPO = "lightx2v/Qwen-Image-Lightning"
|
147 |
LIGHTNING_LORA_WEIGHT = "Qwen-Image-Lightning-8steps-V1.0.safetensors"
|
148 |
|
149 |
-
MAX_SEED =
|
150 |
|
151 |
-
class
|
152 |
-
def __init__(self,
|
153 |
-
self.
|
154 |
|
155 |
def __enter__(self):
|
156 |
self.start_time = time.time()
|
@@ -159,8 +176,8 @@ class Timer:
|
|
159 |
def __exit__(self, exc_type, exc_value, traceback):
|
160 |
self.end_time = time.time()
|
161 |
self.elapsed_time = self.end_time - self.start_time
|
162 |
-
if self.
|
163 |
-
print(f"Elapsed time for {self.
|
164 |
else:
|
165 |
print(f"Elapsed time: {self.elapsed_time:.6f} seconds")
|
166 |
|
@@ -213,26 +230,85 @@ def adjust_generation_mode(speed_mode):
|
|
213 |
return gr.update(value="Base mode selected - 48 steps for best quality"), 48, 4.0
|
214 |
|
215 |
@spaces.GPU(duration=100)
|
216 |
-
def
|
|
|
217 |
pipe.to("cuda")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
218 |
generator = torch.Generator(device="cuda").manual_seed(seed)
|
219 |
-
|
220 |
-
|
221 |
-
|
222 |
-
|
223 |
-
|
224 |
-
|
225 |
-
|
226 |
-
|
227 |
-
|
228 |
-
|
229 |
-
|
230 |
-
|
231 |
-
|
232 |
-
|
|
|
233 |
|
234 |
@spaces.GPU(duration=100)
|
235 |
-
def process_adapter_generation(prompt, cfg_scale, steps, selected_index, randomize_seed, seed, aspect_ratio, lora_scale, speed_mode, progress=gr.Progress(track_tqdm=True)):
|
236 |
if selected_index is None:
|
237 |
raise gr.Error("You must select a LoRA before proceeding.")
|
238 |
|
@@ -253,14 +329,16 @@ def process_adapter_generation(prompt, cfg_scale, steps, selected_index, randomi
|
|
253 |
prompt_mash = prompt
|
254 |
|
255 |
# Always unload any existing LoRAs first to avoid conflicts
|
256 |
-
with
|
257 |
pipe.unload_lora_weights()
|
|
|
|
|
|
|
258 |
|
259 |
-
# Load LoRAs based on speed mode
|
260 |
if speed_mode == "Fast (8 steps)":
|
261 |
-
with
|
262 |
# Load Lightning LoRA first
|
263 |
-
|
264 |
LIGHTNING_LORA_REPO,
|
265 |
weight_name=LIGHTNING_LORA_WEIGHT,
|
266 |
adapter_name="lightning"
|
@@ -268,7 +346,7 @@ def process_adapter_generation(prompt, cfg_scale, steps, selected_index, randomi
|
|
268 |
|
269 |
# Load the selected style LoRA
|
270 |
weight_name = selected_lora.get("weights", None)
|
271 |
-
|
272 |
lora_path,
|
273 |
weight_name=weight_name,
|
274 |
low_cpu_mem_usage=True,
|
@@ -276,29 +354,36 @@ def process_adapter_generation(prompt, cfg_scale, steps, selected_index, randomi
|
|
276 |
)
|
277 |
|
278 |
# Set both adapters active with their weights
|
279 |
-
|
280 |
else:
|
281 |
# Quality mode - only load the style LoRA
|
282 |
-
with
|
283 |
weight_name = selected_lora.get("weights", None)
|
284 |
-
|
285 |
lora_path,
|
286 |
weight_name=weight_name,
|
287 |
low_cpu_mem_usage=True
|
288 |
)
|
289 |
|
290 |
# Set random seed for reproducibility
|
291 |
-
with
|
292 |
if randomize_seed:
|
293 |
seed = random.randint(0, MAX_SEED)
|
294 |
|
295 |
# Get image dimensions from aspect ratio
|
296 |
width, height = compute_image_dimensions(aspect_ratio)
|
297 |
|
298 |
-
|
299 |
-
|
|
|
|
|
|
|
300 |
|
301 |
-
|
|
|
|
|
|
|
|
|
302 |
|
303 |
def fetch_hf_adapter_files(link):
|
304 |
split_link = link.split("/")
|
@@ -422,8 +507,6 @@ def incorporate_custom_adapter(custom_lora):
|
|
422 |
def discard_custom_adapter():
|
423 |
return gr.update(visible=False), gr.update(visible=False), gr.update(), "", None, ""
|
424 |
|
425 |
-
process_adapter_generation.zerogpu = True
|
426 |
-
|
427 |
css = '''
|
428 |
#gen_btn{height: 100%}
|
429 |
#gen_column{align-self: stretch}
|
@@ -436,6 +519,10 @@ css = '''
|
|
436 |
.card_internal img{margin-right: 1em}
|
437 |
.styler{--form-gap-width: 0px !important}
|
438 |
#speed_status{padding: .5em; border-radius: 5px; margin: 1em 0}
|
|
|
|
|
|
|
|
|
439 |
'''
|
440 |
|
441 |
with gr.Blocks(theme="bethecloud/storj_theme", css=css, delete_cache=(120, 120)) as app:
|
@@ -467,6 +554,7 @@ with gr.Blocks(theme="bethecloud/storj_theme", css=css, delete_cache=(120, 120))
|
|
467 |
|
468 |
with gr.Column():
|
469 |
result = gr.Image(label="Generated Image")
|
|
|
470 |
|
471 |
with gr.Row():
|
472 |
aspect_ratio = gr.Dropdown(
|
@@ -508,6 +596,10 @@ with gr.Blocks(theme="bethecloud/storj_theme", css=css, delete_cache=(120, 120))
|
|
508 |
randomize_seed = gr.Checkbox(True, label="Randomize seed")
|
509 |
seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0, randomize=True)
|
510 |
lora_scale = gr.Slider(label="LoRA Scale", minimum=0, maximum=2, step=0.01, value=1.0)
|
|
|
|
|
|
|
|
|
511 |
|
512 |
# Event handlers
|
513 |
gallery.select(
|
@@ -536,8 +628,8 @@ with gr.Blocks(theme="bethecloud/storj_theme", css=css, delete_cache=(120, 120))
|
|
536 |
gr.on(
|
537 |
triggers=[generate_button.click, prompt.submit],
|
538 |
fn=process_adapter_generation,
|
539 |
-
inputs=[prompt, cfg_scale, steps, selected_index, randomize_seed, seed, aspect_ratio, lora_scale, speed_mode],
|
540 |
-
outputs=[result, seed]
|
541 |
)
|
542 |
|
543 |
app.queue()
|
|
|
13 |
import spaces
|
14 |
from diffusers import (
|
15 |
DiffusionPipeline,
|
16 |
+
FlowMatchEulerDiscreteScheduler,
|
17 |
+
AutoencoderKL,
|
18 |
+
AutoPipelineForImage2Image)
|
19 |
from huggingface_hub import (
|
20 |
hf_hub_download,
|
21 |
HfFileSystem,
|
|
|
144 |
base_model, scheduler=scheduler, torch_dtype=dtype
|
145 |
).to(device)
|
146 |
|
147 |
+
taef1 = AutoencoderTiny.from_pretrained("madebyollin/taef1", torch_dtype=dtype).to(device)
|
148 |
+
good_vae = AutoencoderKL.from_pretrained(base_model, subfolder="vae", torch_dtype=dtype).to(device)
|
149 |
+
pipe.vae = taef1
|
150 |
+
pipe_i2i = AutoPipelineForImage2Image.from_pretrained(
|
151 |
+
base_model,
|
152 |
+
vae=good_vae,
|
153 |
+
transformer=pipe.transformer,
|
154 |
+
text_encoder=pipe.text_encoder,
|
155 |
+
tokenizer=pipe.tokenizer,
|
156 |
+
text_encoder_2=pipe.text_encoder_2,
|
157 |
+
tokenizer_2=pipe.tokenizer_2,
|
158 |
+
scheduler=scheduler,
|
159 |
+
torch_dtype=dtype
|
160 |
+
).to(device)
|
161 |
+
|
162 |
# Lightning LoRA info (no global state)
|
163 |
LIGHTNING_LORA_REPO = "lightx2v/Qwen-Image-Lightning"
|
164 |
LIGHTNING_LORA_WEIGHT = "Qwen-Image-Lightning-8steps-V1.0.safetensors"
|
165 |
|
166 |
+
MAX_SEED = 2**32 - 1
|
167 |
|
168 |
+
class calculateDuration:
|
169 |
+
def __init__(self, activity_name=""):
|
170 |
+
self.activity_name = activity_name
|
171 |
|
172 |
def __enter__(self):
|
173 |
self.start_time = time.time()
|
|
|
176 |
def __exit__(self, exc_type, exc_value, traceback):
|
177 |
self.end_time = time.time()
|
178 |
self.elapsed_time = self.end_time - self.start_time
|
179 |
+
if self.activity_name:
|
180 |
+
print(f"Elapsed time for {self.activity_name}: {self.elapsed_time:.6f} seconds")
|
181 |
else:
|
182 |
print(f"Elapsed time: {self.elapsed_time:.6f} seconds")
|
183 |
|
|
|
230 |
return gr.update(value="Base mode selected - 48 steps for best quality"), 48, 4.0
|
231 |
|
232 |
@spaces.GPU(duration=100)
|
233 |
+
def generate_image(prompt_mash, steps, seed, cfg_scale, width, height, lora_scale, negative_prompt=""):
|
234 |
+
generator = torch.Generator(device="cuda").manual_seed(seed)
|
235 |
pipe.to("cuda")
|
236 |
+
|
237 |
+
batch_size = 1
|
238 |
+
prompt = prompt_mash
|
239 |
+
do_classifier_free_guidance = cfg_scale > 1.0
|
240 |
+
prompt_embeds, pooled_prompt_embeds = pipe.encode_prompt(
|
241 |
+
prompt,
|
242 |
+
num_images_per_prompt=1,
|
243 |
+
do_classifier_free_guidance=do_classifier_free_guidance,
|
244 |
+
prompt_2=None,
|
245 |
+
max_sequence_length=256,
|
246 |
+
)
|
247 |
+
height, width = height - height % 16, width - width % 16
|
248 |
+
latents = pipe.prepare_latents(
|
249 |
+
batch_size,
|
250 |
+
pipe.transformer.config.in_channels,
|
251 |
+
height,
|
252 |
+
width,
|
253 |
+
dtype,
|
254 |
+
device,
|
255 |
+
generator,
|
256 |
+
latents=None,
|
257 |
+
)
|
258 |
+
pipe.scheduler.set_timesteps(steps)
|
259 |
+
timesteps = pipe.scheduler.timesteps
|
260 |
+
joint_attention_kwargs = {"scale": lora_scale}
|
261 |
+
for i in range(steps):
|
262 |
+
t = pipe.scheduler.sigmas[i]
|
263 |
+
latent_model_input = latents
|
264 |
+
with torch.no_grad():
|
265 |
+
noise_pred = pipe.transformer(
|
266 |
+
hidden_states=latent_model_input,
|
267 |
+
timestep=t,
|
268 |
+
guidance=cfg_scale,
|
269 |
+
pooled_projections=pooled_prompt_embeds,
|
270 |
+
encoder_hidden_states=prompt_embeds,
|
271 |
+
joint_attention_kwargs=joint_attention_kwargs,
|
272 |
+
return_dict=False,
|
273 |
+
)[0]
|
274 |
+
latents = pipe.scheduler.step(
|
275 |
+
model_output=noise_pred,
|
276 |
+
timestep=t,
|
277 |
+
sample=latent_model_input,
|
278 |
+
return_dict=False,
|
279 |
+
)[0]
|
280 |
+
# preview
|
281 |
+
with torch.no_grad():
|
282 |
+
decoded = pipe.vae.decode(latents / pipe.vae.config.scaling_factor, return_dict=False)[0]
|
283 |
+
image = pipe.image_processor.pt_to_pil(decoded)[0]
|
284 |
+
yield image
|
285 |
+
# final
|
286 |
+
with torch.no_grad():
|
287 |
+
decoded = good_vae.decode(latents / good_vae.config.scaling_factor, return_dict=False)[0]
|
288 |
+
image = pipe.image_processor.pt_to_pil(decoded)[0]
|
289 |
+
yield image
|
290 |
+
|
291 |
+
@spaces.GPU(duration=100)
|
292 |
+
def generate_image_to_image(prompt_mash, image_input_path, image_strength, steps, cfg_scale, width, height, lora_scale, seed):
|
293 |
generator = torch.Generator(device="cuda").manual_seed(seed)
|
294 |
+
pipe_i2i.to("cuda")
|
295 |
+
image_input = load_image(image_input_path)
|
296 |
+
final_image = pipe_i2i(
|
297 |
+
prompt=prompt_mash,
|
298 |
+
image=image_input,
|
299 |
+
strength=image_strength,
|
300 |
+
num_inference_steps=steps,
|
301 |
+
guidance_scale=cfg_scale,
|
302 |
+
width=width,
|
303 |
+
height=height,
|
304 |
+
generator=generator,
|
305 |
+
joint_attention_kwargs={"scale": lora_scale},
|
306 |
+
output_type="pil",
|
307 |
+
).images[0]
|
308 |
+
return final_image
|
309 |
|
310 |
@spaces.GPU(duration=100)
|
311 |
+
def process_adapter_generation(prompt, cfg_scale, steps, selected_index, randomize_seed, seed, aspect_ratio, lora_scale, speed_mode, image_input, image_strength, negative_prompt="", progress=gr.Progress(track_tqdm=True)):
|
312 |
if selected_index is None:
|
313 |
raise gr.Error("You must select a LoRA before proceeding.")
|
314 |
|
|
|
329 |
prompt_mash = prompt
|
330 |
|
331 |
# Always unload any existing LoRAs first to avoid conflicts
|
332 |
+
with calculateDuration("Unloading existing LoRAs"):
|
333 |
pipe.unload_lora_weights()
|
334 |
+
pipe_i2i.unload_lora_weights()
|
335 |
+
|
336 |
+
pipe_to_use = pipe_i2i if image_input is not None else pipe
|
337 |
|
|
|
338 |
if speed_mode == "Fast (8 steps)":
|
339 |
+
with calculateDuration("Loading Lightning LoRA and style LoRA"):
|
340 |
# Load Lightning LoRA first
|
341 |
+
pipe_to_use.load_lora_weights(
|
342 |
LIGHTNING_LORA_REPO,
|
343 |
weight_name=LIGHTNING_LORA_WEIGHT,
|
344 |
adapter_name="lightning"
|
|
|
346 |
|
347 |
# Load the selected style LoRA
|
348 |
weight_name = selected_lora.get("weights", None)
|
349 |
+
pipe_to_use.load_lora_weights(
|
350 |
lora_path,
|
351 |
weight_name=weight_name,
|
352 |
low_cpu_mem_usage=True,
|
|
|
354 |
)
|
355 |
|
356 |
# Set both adapters active with their weights
|
357 |
+
pipe_to_use.set_adapters(["lightning", "style"], adapter_weights=[1.0, lora_scale])
|
358 |
else:
|
359 |
# Quality mode - only load the style LoRA
|
360 |
+
with calculateDuration(f"Loading LoRA weights for {selected_lora['title']}"):
|
361 |
weight_name = selected_lora.get("weights", None)
|
362 |
+
pipe_to_use.load_lora_weights(
|
363 |
lora_path,
|
364 |
weight_name=weight_name,
|
365 |
low_cpu_mem_usage=True
|
366 |
)
|
367 |
|
368 |
# Set random seed for reproducibility
|
369 |
+
with calculateDuration("Randomizing seed"):
|
370 |
if randomize_seed:
|
371 |
seed = random.randint(0, MAX_SEED)
|
372 |
|
373 |
# Get image dimensions from aspect ratio
|
374 |
width, height = compute_image_dimensions(aspect_ratio)
|
375 |
|
376 |
+
if image_input is not None:
|
377 |
+
final_image = generate_image_to_image(prompt_mash, image_input, image_strength, steps, cfg_scale, width, height, lora_scale, seed)
|
378 |
+
yield final_image, seed, gr.update(visible=False)
|
379 |
+
else:
|
380 |
+
image_generator = generate_image(prompt_mash, steps, seed, cfg_scale, width, height, lora_scale, negative_prompt)
|
381 |
|
382 |
+
step_counter = 0
|
383 |
+
for image in image_generator:
|
384 |
+
step_counter += 1
|
385 |
+
progress_bar = f'<div class="progress-container"><div class="progress-bar" style="--current: {step_counter}; --total: {steps};"></div></div>'
|
386 |
+
yield image, seed, gr.update(value=progress_bar, visible=True)
|
387 |
|
388 |
def fetch_hf_adapter_files(link):
|
389 |
split_link = link.split("/")
|
|
|
507 |
def discard_custom_adapter():
|
508 |
return gr.update(visible=False), gr.update(visible=False), gr.update(), "", None, ""
|
509 |
|
|
|
|
|
510 |
css = '''
|
511 |
#gen_btn{height: 100%}
|
512 |
#gen_column{align-self: stretch}
|
|
|
519 |
.card_internal img{margin-right: 1em}
|
520 |
.styler{--form-gap-width: 0px !important}
|
521 |
#speed_status{padding: .5em; border-radius: 5px; margin: 1em 0}
|
522 |
+
#progress{height:30px}
|
523 |
+
#progress .generating{display:none}
|
524 |
+
.progress-container {width: 100%;height: 30px;background-color: #f0f0f0;border-radius: 15px;overflow: hidden;margin-bottom: 20px}
|
525 |
+
.progress-bar {height: 100%;background-color: #4f46e5;width: calc(var(--current) / var(--total) * 100%);transition: width 0.5s ease-in-out}
|
526 |
'''
|
527 |
|
528 |
with gr.Blocks(theme="bethecloud/storj_theme", css=css, delete_cache=(120, 120)) as app:
|
|
|
554 |
|
555 |
with gr.Column():
|
556 |
result = gr.Image(label="Generated Image")
|
557 |
+
progress_html = gr.HTML(visible=False, elem_id="progress")
|
558 |
|
559 |
with gr.Row():
|
560 |
aspect_ratio = gr.Dropdown(
|
|
|
596 |
randomize_seed = gr.Checkbox(True, label="Randomize seed")
|
597 |
seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0, randomize=True)
|
598 |
lora_scale = gr.Slider(label="LoRA Scale", minimum=0, maximum=2, step=0.01, value=1.0)
|
599 |
+
|
600 |
+
with gr.Row():
|
601 |
+
image_input = gr.Image(label="Input Image for Image2Image", type="filepath")
|
602 |
+
image_strength = gr.Slider(label="Image Strength", minimum=0, maximum=1, step=0.01, value=0.35)
|
603 |
|
604 |
# Event handlers
|
605 |
gallery.select(
|
|
|
628 |
gr.on(
|
629 |
triggers=[generate_button.click, prompt.submit],
|
630 |
fn=process_adapter_generation,
|
631 |
+
inputs=[prompt, cfg_scale, steps, selected_index, randomize_seed, seed, aspect_ratio, lora_scale, speed_mode, image_input, image_strength],
|
632 |
+
outputs=[result, seed, progress_html]
|
633 |
)
|
634 |
|
635 |
app.queue()
|