Update app.py
Browse files
app.py
CHANGED
|
@@ -23,6 +23,9 @@ device = "cuda" if torch.cuda.is_available() else "cpu"
|
|
| 23 |
MAX_SEED = np.iinfo(np.int32).max
|
| 24 |
MAX_IMAGE_SIZE = 1024
|
| 25 |
|
|
|
|
|
|
|
|
|
|
| 26 |
hf_client = InferenceClient(
|
| 27 |
api_key=os.environ.get("HF_TOKEN"),
|
| 28 |
)
|
|
@@ -78,6 +81,13 @@ pipe = Flux2Pipeline.from_pretrained(
|
|
| 78 |
transformer=dit,
|
| 79 |
torch_dtype=torch.bfloat16
|
| 80 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 81 |
pipe.to(device)
|
| 82 |
|
| 83 |
# Pull pre-compiled Flux2 Transformer blocks from HF hub
|
|
@@ -157,14 +167,17 @@ def update_dimensions_from_image(image_list):
|
|
| 157 |
|
| 158 |
return new_width, new_height
|
| 159 |
|
| 160 |
-
# Updated duration function
|
| 161 |
-
def get_duration(prompt_embeds, image_list, width, height, num_inference_steps, guidance_scale, seed, progress=gr.Progress(track_tqdm=True)):
|
| 162 |
num_images = 0 if image_list is None else len(image_list)
|
| 163 |
step_duration = 1 + 0.8 * num_images
|
|
|
|
|
|
|
|
|
|
| 164 |
return max(65, num_inference_steps * step_duration + 10)
|
| 165 |
|
| 166 |
@spaces.GPU(duration=get_duration)
|
| 167 |
-
def generate_image(prompt_embeds, image_list, width, height, num_inference_steps, guidance_scale, seed, progress=gr.Progress(track_tqdm=True)):
|
| 168 |
# Move embeddings to GPU only when inside the GPU decorated function
|
| 169 |
prompt_embeds = prompt_embeds.to(device)
|
| 170 |
|
|
@@ -173,13 +186,19 @@ def generate_image(prompt_embeds, image_list, width, height, num_inference_steps
|
|
| 173 |
pipe_kwargs = {
|
| 174 |
"prompt_embeds": prompt_embeds,
|
| 175 |
"image": image_list,
|
| 176 |
-
"num_inference_steps": num_inference_steps,
|
| 177 |
"guidance_scale": guidance_scale,
|
| 178 |
"generator": generator,
|
| 179 |
"width": width,
|
| 180 |
"height": height,
|
| 181 |
}
|
| 182 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 183 |
# Progress bar for the actual generation steps
|
| 184 |
if progress:
|
| 185 |
progress(0, desc="Starting generation...")
|
|
@@ -187,7 +206,7 @@ def generate_image(prompt_embeds, image_list, width, height, num_inference_steps
|
|
| 187 |
image = pipe(**pipe_kwargs).images[0]
|
| 188 |
return image
|
| 189 |
|
| 190 |
-
def infer(prompt, input_images=None, seed=42, randomize_seed=False, width=1024, height=1024, num_inference_steps=
|
| 191 |
|
| 192 |
if randomize_seed:
|
| 193 |
seed = random.randint(0, MAX_SEED)
|
|
@@ -221,7 +240,8 @@ def infer(prompt, input_images=None, seed=42, randomize_seed=False, width=1024,
|
|
| 221 |
height,
|
| 222 |
num_inference_steps,
|
| 223 |
guidance_scale,
|
| 224 |
-
seed,
|
|
|
|
| 225 |
progress
|
| 226 |
)
|
| 227 |
|
|
@@ -252,8 +272,8 @@ css="""
|
|
| 252 |
with gr.Blocks() as demo:
|
| 253 |
|
| 254 |
with gr.Column(elem_id="col-container"):
|
| 255 |
-
gr.Markdown(f"""# FLUX.2 [dev]
|
| 256 |
-
FLUX.2 [dev]
|
| 257 |
""")
|
| 258 |
with gr.Row():
|
| 259 |
with gr.Column():
|
|
@@ -278,6 +298,12 @@ FLUX.2 [dev] is a 32B model rectified flow capable of generating, editing and co
|
|
| 278 |
)
|
| 279 |
|
| 280 |
with gr.Accordion("Advanced Settings", open=False):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 281 |
prompt_upsampling = gr.Checkbox(
|
| 282 |
label="Prompt Upsampling",
|
| 283 |
value=True,
|
|
@@ -315,7 +341,7 @@ FLUX.2 [dev] is a 32B model rectified flow capable of generating, editing and co
|
|
| 315 |
with gr.Row():
|
| 316 |
|
| 317 |
num_inference_steps = gr.Slider(
|
| 318 |
-
label="Number of inference steps",
|
| 319 |
minimum=1,
|
| 320 |
maximum=100,
|
| 321 |
step=1,
|
|
@@ -327,7 +353,7 @@ FLUX.2 [dev] is a 32B model rectified flow capable of generating, editing and co
|
|
| 327 |
minimum=0.0,
|
| 328 |
maximum=10.0,
|
| 329 |
step=0.1,
|
| 330 |
-
value=
|
| 331 |
)
|
| 332 |
|
| 333 |
|
|
@@ -363,7 +389,7 @@ FLUX.2 [dev] is a 32B model rectified flow capable of generating, editing and co
|
|
| 363 |
gr.on(
|
| 364 |
triggers=[run_button.click, prompt.submit],
|
| 365 |
fn=infer,
|
| 366 |
-
inputs=[prompt, input_images, seed, randomize_seed, width, height, num_inference_steps, guidance_scale, prompt_upsampling],
|
| 367 |
outputs=[result, seed]
|
| 368 |
)
|
| 369 |
|
|
|
|
| 23 |
MAX_SEED = np.iinfo(np.int32).max
|
| 24 |
MAX_IMAGE_SIZE = 1024
|
| 25 |
|
| 26 |
+
# Pre-shifted custom sigmas for 8-step turbo inference
|
| 27 |
+
TURBO_SIGMAS = [1.0, 0.6509, 0.4374, 0.2932, 0.1893, 0.1108, 0.0495, 0.00031]
|
| 28 |
+
|
| 29 |
hf_client = InferenceClient(
|
| 30 |
api_key=os.environ.get("HF_TOKEN"),
|
| 31 |
)
|
|
|
|
| 81 |
transformer=dit,
|
| 82 |
torch_dtype=torch.bfloat16
|
| 83 |
)
|
| 84 |
+
|
| 85 |
+
# Load the Turbo LoRA
|
| 86 |
+
pipe.load_lora_weights(
|
| 87 |
+
"fal/FLUX.2-Turbo",
|
| 88 |
+
weight_name="flux.2-turbo-lora.safetensors"
|
| 89 |
+
)
|
| 90 |
+
|
| 91 |
pipe.to(device)
|
| 92 |
|
| 93 |
# Pull pre-compiled Flux2 Transformer blocks from HF hub
|
|
|
|
| 167 |
|
| 168 |
return new_width, new_height
|
| 169 |
|
| 170 |
+
# Updated duration function for Turbo (much faster with fewer steps)
|
| 171 |
+
def get_duration(prompt_embeds, image_list, width, height, num_inference_steps, guidance_scale, seed, use_turbo, progress=gr.Progress(track_tqdm=True)):
|
| 172 |
num_images = 0 if image_list is None else len(image_list)
|
| 173 |
step_duration = 1 + 0.8 * num_images
|
| 174 |
+
# Turbo mode uses fewer steps, so shorter duration
|
| 175 |
+
if use_turbo:
|
| 176 |
+
return max(30, 8 * step_duration + 10) # Fixed 8 steps for turbo
|
| 177 |
return max(65, num_inference_steps * step_duration + 10)
|
| 178 |
|
| 179 |
@spaces.GPU(duration=get_duration)
|
| 180 |
+
def generate_image(prompt_embeds, image_list, width, height, num_inference_steps, guidance_scale, seed, use_turbo, progress=gr.Progress(track_tqdm=True)):
|
| 181 |
# Move embeddings to GPU only when inside the GPU decorated function
|
| 182 |
prompt_embeds = prompt_embeds.to(device)
|
| 183 |
|
|
|
|
| 186 |
pipe_kwargs = {
|
| 187 |
"prompt_embeds": prompt_embeds,
|
| 188 |
"image": image_list,
|
|
|
|
| 189 |
"guidance_scale": guidance_scale,
|
| 190 |
"generator": generator,
|
| 191 |
"width": width,
|
| 192 |
"height": height,
|
| 193 |
}
|
| 194 |
|
| 195 |
+
# Use Turbo sigmas or regular inference steps
|
| 196 |
+
if use_turbo:
|
| 197 |
+
pipe_kwargs["sigmas"] = TURBO_SIGMAS
|
| 198 |
+
pipe_kwargs["num_inference_steps"] = 8 # Turbo always uses 8 steps
|
| 199 |
+
else:
|
| 200 |
+
pipe_kwargs["num_inference_steps"] = num_inference_steps
|
| 201 |
+
|
| 202 |
# Progress bar for the actual generation steps
|
| 203 |
if progress:
|
| 204 |
progress(0, desc="Starting generation...")
|
|
|
|
| 206 |
image = pipe(**pipe_kwargs).images[0]
|
| 207 |
return image
|
| 208 |
|
| 209 |
+
def infer(prompt, input_images=None, seed=42, randomize_seed=False, width=1024, height=1024, num_inference_steps=8, guidance_scale=2.5, prompt_upsampling=False, use_turbo=True, progress=gr.Progress(track_tqdm=True)):
|
| 210 |
|
| 211 |
if randomize_seed:
|
| 212 |
seed = random.randint(0, MAX_SEED)
|
|
|
|
| 240 |
height,
|
| 241 |
num_inference_steps,
|
| 242 |
guidance_scale,
|
| 243 |
+
seed,
|
| 244 |
+
use_turbo,
|
| 245 |
progress
|
| 246 |
)
|
| 247 |
|
|
|
|
| 272 |
with gr.Blocks() as demo:
|
| 273 |
|
| 274 |
with gr.Column(elem_id="col-container"):
|
| 275 |
+
gr.Markdown(f"""# FLUX.2 [dev] Turbo
|
| 276 |
+
FLUX.2 [dev] with [Turbo LoRA by fal](https://huggingface.co/fal/FLUX.2-Turbo) - a 32B rectified flow model capable of generating, editing and combining images based on text instructions in just 8 steps [[model](https://huggingface.co/black-forest-labs/FLUX.2-dev)], [[blog](https://bfl.ai/blog/flux-2)]
|
| 277 |
""")
|
| 278 |
with gr.Row():
|
| 279 |
with gr.Column():
|
|
|
|
| 298 |
)
|
| 299 |
|
| 300 |
with gr.Accordion("Advanced Settings", open=False):
|
| 301 |
+
use_turbo = gr.Checkbox(
|
| 302 |
+
label="Use Turbo Mode (8 steps)",
|
| 303 |
+
value=True,
|
| 304 |
+
info="Enable Turbo LoRA for fast 8-step generation"
|
| 305 |
+
)
|
| 306 |
+
|
| 307 |
prompt_upsampling = gr.Checkbox(
|
| 308 |
label="Prompt Upsampling",
|
| 309 |
value=True,
|
|
|
|
| 341 |
with gr.Row():
|
| 342 |
|
| 343 |
num_inference_steps = gr.Slider(
|
| 344 |
+
label="Number of inference steps (ignored in Turbo mode)",
|
| 345 |
minimum=1,
|
| 346 |
maximum=100,
|
| 347 |
step=1,
|
|
|
|
| 353 |
minimum=0.0,
|
| 354 |
maximum=10.0,
|
| 355 |
step=0.1,
|
| 356 |
+
value=2.5,
|
| 357 |
)
|
| 358 |
|
| 359 |
|
|
|
|
| 389 |
gr.on(
|
| 390 |
triggers=[run_button.click, prompt.submit],
|
| 391 |
fn=infer,
|
| 392 |
+
inputs=[prompt, input_images, seed, randomize_seed, width, height, num_inference_steps, guidance_scale, prompt_upsampling, use_turbo],
|
| 393 |
outputs=[result, seed]
|
| 394 |
)
|
| 395 |
|