Spaces:
Runtime error
Runtime error
Commit
•
920f544
1
Parent(s):
9efde7c
Create app.py
Browse files
app.py
ADDED
@@ -0,0 +1,362 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import gc
|
3 |
+
import gradio as gr
|
4 |
+
import numpy as np
|
5 |
+
import torch
|
6 |
+
import json
|
7 |
+
import spaces
|
8 |
+
import config
|
9 |
+
import utils
|
10 |
+
import logging
|
11 |
+
from PIL import Image, PngImagePlugin
|
12 |
+
from datetime import datetime
|
13 |
+
from diffusers.models import AutoencoderKL
|
14 |
+
from diffusers import StableDiffusionXLPipeline, StableDiffusionXLImg2ImgPipeline
|
15 |
+
|
16 |
+
logging.basicConfig(level=logging.INFO)
|
17 |
+
logger = logging.getLogger(__name__)
|
18 |
+
|
19 |
+
DESCRIPTION = "Animagine XL 3.0"
|
20 |
+
if not torch.cuda.is_available():
|
21 |
+
DESCRIPTION += "\n<p>Running on CPU 🥶 This demo does not work on CPU. </p>"
|
22 |
+
IS_COLAB = utils.is_google_colab() or os.getenv("IS_COLAB") == "1"
|
23 |
+
HF_TOKEN = os.getenv("HF_TOKEN")
|
24 |
+
CACHE_EXAMPLES = torch.cuda.is_available() and os.getenv("CACHE_EXAMPLES") == "1"
|
25 |
+
MIN_IMAGE_SIZE = int(os.getenv("MIN_IMAGE_SIZE", "512"))
|
26 |
+
MAX_IMAGE_SIZE = int(os.getenv("MAX_IMAGE_SIZE", "2048"))
|
27 |
+
USE_TORCH_COMPILE = os.getenv("USE_TORCH_COMPILE") == "1"
|
28 |
+
ENABLE_CPU_OFFLOAD = os.getenv("ENABLE_CPU_OFFLOAD") == "1"
|
29 |
+
OUTPUT_DIR = os.getenv("OUTPUT_DIR", "./outputs")
|
30 |
+
|
31 |
+
MODEL = os.getenv(
|
32 |
+
"MODEL",
|
33 |
+
"https://huggingface.co/cagliostrolab/animagine-xl-3.0/blob/main/animagine-xl-3.0.safetensors",
|
34 |
+
)
|
35 |
+
|
36 |
+
torch.backends.cudnn.deterministic = True
|
37 |
+
torch.backends.cudnn.benchmark = False
|
38 |
+
|
39 |
+
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
|
40 |
+
|
41 |
+
|
42 |
+
def load_pipeline(model_name):
|
43 |
+
vae = AutoencoderKL.from_pretrained(
|
44 |
+
"madebyollin/sdxl-vae-fp16-fix",
|
45 |
+
torch_dtype=torch.float16,
|
46 |
+
)
|
47 |
+
pipeline = (
|
48 |
+
StableDiffusionXLPipeline.from_single_file
|
49 |
+
if MODEL.endswith(".safetensors")
|
50 |
+
else StableDiffusionXLPipeline.from_pretrained
|
51 |
+
)
|
52 |
+
|
53 |
+
pipe = pipeline(
|
54 |
+
model_name,
|
55 |
+
vae=vae,
|
56 |
+
torch_dtype=torch.float16,
|
57 |
+
custom_pipeline="lpw_stable_diffusion_xl",
|
58 |
+
use_safetensors=True,
|
59 |
+
add_watermarker=False,
|
60 |
+
use_auth_token=HF_TOKEN,
|
61 |
+
variant="fp16",
|
62 |
+
)
|
63 |
+
|
64 |
+
pipe.to(device)
|
65 |
+
return pipe
|
66 |
+
|
67 |
+
|
68 |
+
@spaces.GPU
|
69 |
+
def generate(
|
70 |
+
prompt: str,
|
71 |
+
negative_prompt: str = "",
|
72 |
+
seed: int = 0,
|
73 |
+
custom_width: int = 1024,
|
74 |
+
custom_height: int = 1024,
|
75 |
+
guidance_scale: float = 7.0,
|
76 |
+
num_inference_steps: int = 28,
|
77 |
+
sampler: str = "Euler a",
|
78 |
+
aspect_ratio_selector: str = "896 x 1152",
|
79 |
+
use_upscaler: bool = False,
|
80 |
+
upscaler_strength: float = 0.55,
|
81 |
+
upscale_by: float = 1.5,
|
82 |
+
progress=gr.Progress(track_tqdm=True),
|
83 |
+
) -> Image:
|
84 |
+
generator = utils.seed_everything(seed)
|
85 |
+
|
86 |
+
width, height = utils.aspect_ratio_handler(
|
87 |
+
aspect_ratio_selector,
|
88 |
+
custom_width,
|
89 |
+
custom_height,
|
90 |
+
)
|
91 |
+
|
92 |
+
width, height = utils.preprocess_image_dimensions(width, height)
|
93 |
+
|
94 |
+
backup_scheduler = pipe.scheduler
|
95 |
+
pipe.scheduler = utils.get_scheduler(pipe.scheduler.config, sampler)
|
96 |
+
|
97 |
+
if use_upscaler:
|
98 |
+
upscaler_pipe = StableDiffusionXLImg2ImgPipeline(**pipe.components)
|
99 |
+
metadata = {
|
100 |
+
"prompt": prompt,
|
101 |
+
"negative_prompt": negative_prompt,
|
102 |
+
"resolution": f"{width} x {height}",
|
103 |
+
"guidance_scale": guidance_scale,
|
104 |
+
"num_inference_steps": num_inference_steps,
|
105 |
+
"seed": seed,
|
106 |
+
"sampler": sampler,
|
107 |
+
}
|
108 |
+
|
109 |
+
if use_upscaler:
|
110 |
+
new_width = int(width * upscale_by)
|
111 |
+
new_height = int(height * upscale_by)
|
112 |
+
metadata["use_upscaler"] = {
|
113 |
+
"upscale_method": "nearest-exact",
|
114 |
+
"upscaler_strength": upscaler_strength,
|
115 |
+
"upscale_by": upscale_by,
|
116 |
+
"new_resolution": f"{new_width} x {new_height}",
|
117 |
+
}
|
118 |
+
else:
|
119 |
+
metadata["use_upscaler"] = None
|
120 |
+
logger.info(json.dumps(metadata, indent=4))
|
121 |
+
|
122 |
+
try:
|
123 |
+
if use_upscaler:
|
124 |
+
latents = pipe(
|
125 |
+
prompt=prompt,
|
126 |
+
negative_prompt=negative_prompt,
|
127 |
+
width=width,
|
128 |
+
height=height,
|
129 |
+
guidance_scale=guidance_scale,
|
130 |
+
num_inference_steps=num_inference_steps,
|
131 |
+
generator=generator,
|
132 |
+
output_type="latent",
|
133 |
+
).images
|
134 |
+
upscaled_latents = utils.upscale(latents, "nearest-exact", upscale_by)
|
135 |
+
images = upscaler_pipe(
|
136 |
+
prompt=prompt,
|
137 |
+
negative_prompt=negative_prompt,
|
138 |
+
image=upscaled_latents,
|
139 |
+
guidance_scale=guidance_scale,
|
140 |
+
num_inference_steps=num_inference_steps,
|
141 |
+
strength=upscaler_strength,
|
142 |
+
generator=generator,
|
143 |
+
output_type="pil",
|
144 |
+
).images
|
145 |
+
else:
|
146 |
+
images = pipe(
|
147 |
+
prompt=prompt,
|
148 |
+
negative_prompt=negative_prompt,
|
149 |
+
width=width,
|
150 |
+
height=height,
|
151 |
+
guidance_scale=guidance_scale,
|
152 |
+
num_inference_steps=num_inference_steps,
|
153 |
+
generator=generator,
|
154 |
+
output_type="pil",
|
155 |
+
).images
|
156 |
+
|
157 |
+
if images and IS_COLAB:
|
158 |
+
for image in images:
|
159 |
+
filepath = utils.save_image(image, metadata, OUTPUT_DIR)
|
160 |
+
logger.info(f"Image saved as {filepath} with metadata")
|
161 |
+
|
162 |
+
return images, metadata
|
163 |
+
except Exception as e:
|
164 |
+
logger.exception(f"An error occurred: {e}")
|
165 |
+
raise
|
166 |
+
finally:
|
167 |
+
if use_upscaler:
|
168 |
+
del upscaler_pipe
|
169 |
+
pipe.scheduler = backup_scheduler
|
170 |
+
utils.free_memory()
|
171 |
+
|
172 |
+
|
173 |
+
if torch.cuda.is_available():
|
174 |
+
pipe = load_pipeline(MODEL)
|
175 |
+
logger.info("Loaded on Device!")
|
176 |
+
else:
|
177 |
+
pipe = None
|
178 |
+
|
179 |
+
with gr.Blocks(css="style.css") as demo:
|
180 |
+
title = gr.HTML(
|
181 |
+
f"""<h1><span>{DESCRIPTION}</span></h1>""",
|
182 |
+
elem_id="title",
|
183 |
+
)
|
184 |
+
gr.Markdown(
|
185 |
+
f"""Gradio demo for [cagliostrolab/animagine-xl-3.0](https://huggingface.co/cagliostrolab/animagine-xl-3.0)""",
|
186 |
+
elem_id="subtitle",
|
187 |
+
)
|
188 |
+
gr.DuplicateButton(
|
189 |
+
value="Duplicate Space for private use",
|
190 |
+
elem_id="duplicate-button",
|
191 |
+
visible=os.getenv("SHOW_DUPLICATE_BUTTON") == "1",
|
192 |
+
)
|
193 |
+
with gr.Group():
|
194 |
+
with gr.Row():
|
195 |
+
prompt = gr.Text(
|
196 |
+
label="Prompt",
|
197 |
+
show_label=False,
|
198 |
+
max_lines=5,
|
199 |
+
placeholder="Enter your prompt",
|
200 |
+
container=False,
|
201 |
+
)
|
202 |
+
run_button = gr.Button(
|
203 |
+
"Generate",
|
204 |
+
variant="primary",
|
205 |
+
scale=0
|
206 |
+
)
|
207 |
+
result = gr.Gallery(
|
208 |
+
label="Result",
|
209 |
+
columns=1,
|
210 |
+
preview=True,
|
211 |
+
show_label=False
|
212 |
+
)
|
213 |
+
with gr.Accordion(label="Advanced Settings", open=False):
|
214 |
+
negative_prompt = gr.Text(
|
215 |
+
label="Negative Prompt",
|
216 |
+
max_lines=5,
|
217 |
+
placeholder="Enter a negative prompt",
|
218 |
+
)
|
219 |
+
aspect_ratio_selector = gr.Radio(
|
220 |
+
label="Aspect Ratio",
|
221 |
+
choices=config.aspect_ratios,
|
222 |
+
value="896 x 1152",
|
223 |
+
container=True,
|
224 |
+
)
|
225 |
+
with gr.Group(visible=False) as custom_resolution:
|
226 |
+
with gr.Row():
|
227 |
+
custom_width = gr.Slider(
|
228 |
+
label="Width",
|
229 |
+
minimum=MIN_IMAGE_SIZE,
|
230 |
+
maximum=MAX_IMAGE_SIZE,
|
231 |
+
step=8,
|
232 |
+
value=1024,
|
233 |
+
)
|
234 |
+
custom_height = gr.Slider(
|
235 |
+
label="Height",
|
236 |
+
minimum=MIN_IMAGE_SIZE,
|
237 |
+
maximum=MAX_IMAGE_SIZE,
|
238 |
+
step=8,
|
239 |
+
value=1024,
|
240 |
+
)
|
241 |
+
use_upscaler = gr.Checkbox(label="Use Upscaler", value=False)
|
242 |
+
with gr.Row() as upscaler_row:
|
243 |
+
upscaler_strength = gr.Slider(
|
244 |
+
label="Strength",
|
245 |
+
minimum=0,
|
246 |
+
maximum=1,
|
247 |
+
step=0.05,
|
248 |
+
value=0.55,
|
249 |
+
visible=False,
|
250 |
+
)
|
251 |
+
upscale_by = gr.Slider(
|
252 |
+
label="Upscale by",
|
253 |
+
minimum=1,
|
254 |
+
maximum=1.5,
|
255 |
+
step=0.1,
|
256 |
+
value=1.5,
|
257 |
+
visible=False,
|
258 |
+
)
|
259 |
+
|
260 |
+
sampler = gr.Dropdown(
|
261 |
+
label="Sampler",
|
262 |
+
choices=config.sampler_list,
|
263 |
+
interactive=True,
|
264 |
+
value="Euler a",
|
265 |
+
)
|
266 |
+
with gr.Row():
|
267 |
+
seed = gr.Slider(
|
268 |
+
label="Seed", minimum=0, maximum=utils.MAX_SEED, step=1, value=0
|
269 |
+
)
|
270 |
+
randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
|
271 |
+
with gr.Group():
|
272 |
+
with gr.Row():
|
273 |
+
guidance_scale = gr.Slider(
|
274 |
+
label="Guidance scale",
|
275 |
+
minimum=1,
|
276 |
+
maximum=12,
|
277 |
+
step=0.1,
|
278 |
+
value=7.0,
|
279 |
+
)
|
280 |
+
num_inference_steps = gr.Slider(
|
281 |
+
label="Number of inference steps",
|
282 |
+
minimum=1,
|
283 |
+
maximum=50,
|
284 |
+
step=1,
|
285 |
+
value=28,
|
286 |
+
)
|
287 |
+
with gr.Accordion(label="Generation Parameters", open=False):
|
288 |
+
gr_metadata = gr.JSON(label="Metadata", show_label=False)
|
289 |
+
gr.Examples(
|
290 |
+
examples=config.examples,
|
291 |
+
inputs=prompt,
|
292 |
+
outputs=[result, gr_metadata],
|
293 |
+
fn=lambda *args, **kwargs: generate(*args, use_upscaler=True, **kwargs),
|
294 |
+
cache_examples=CACHE_EXAMPLES,
|
295 |
+
)
|
296 |
+
use_upscaler.change(
|
297 |
+
fn=lambda x: [gr.update(visible=x), gr.update(visible=x)],
|
298 |
+
inputs=use_upscaler,
|
299 |
+
outputs=[upscaler_strength, upscale_by],
|
300 |
+
queue=False,
|
301 |
+
api_name=False,
|
302 |
+
)
|
303 |
+
aspect_ratio_selector.change(
|
304 |
+
fn=lambda x: gr.update(visible=x == "Custom"),
|
305 |
+
inputs=aspect_ratio_selector,
|
306 |
+
outputs=custom_resolution,
|
307 |
+
queue=False,
|
308 |
+
api_name=False,
|
309 |
+
)
|
310 |
+
|
311 |
+
inputs = [
|
312 |
+
prompt,
|
313 |
+
negative_prompt,
|
314 |
+
seed,
|
315 |
+
custom_width,
|
316 |
+
custom_height,
|
317 |
+
guidance_scale,
|
318 |
+
num_inference_steps,
|
319 |
+
sampler,
|
320 |
+
aspect_ratio_selector,
|
321 |
+
use_upscaler,
|
322 |
+
upscaler_strength,
|
323 |
+
upscale_by,
|
324 |
+
]
|
325 |
+
|
326 |
+
prompt.submit(
|
327 |
+
fn=utils.randomize_seed_fn,
|
328 |
+
inputs=[seed, randomize_seed],
|
329 |
+
outputs=seed,
|
330 |
+
queue=False,
|
331 |
+
api_name=False,
|
332 |
+
).then(
|
333 |
+
fn=generate,
|
334 |
+
inputs=inputs,
|
335 |
+
outputs=result,
|
336 |
+
api_name="run",
|
337 |
+
)
|
338 |
+
negative_prompt.submit(
|
339 |
+
fn=utils.randomize_seed_fn,
|
340 |
+
inputs=[seed, randomize_seed],
|
341 |
+
outputs=seed,
|
342 |
+
queue=False,
|
343 |
+
api_name=False,
|
344 |
+
).then(
|
345 |
+
fn=generate,
|
346 |
+
inputs=inputs,
|
347 |
+
outputs=result,
|
348 |
+
api_name=False,
|
349 |
+
)
|
350 |
+
run_button.click(
|
351 |
+
fn=utils.randomize_seed_fn,
|
352 |
+
inputs=[seed, randomize_seed],
|
353 |
+
outputs=seed,
|
354 |
+
queue=False,
|
355 |
+
api_name=False,
|
356 |
+
).then(
|
357 |
+
fn=generate,
|
358 |
+
inputs=inputs,
|
359 |
+
outputs=[result, gr_metadata],
|
360 |
+
api_name=False,
|
361 |
+
)
|
362 |
+
demo.queue(max_size=20).launch(debug=IS_COLAB, share=IS_COLAB)
|