Spaces:
Running
on
Zero
Running
on
Zero
Update app.py
Browse files
app.py
CHANGED
|
@@ -2,447 +2,229 @@ import gradio as gr
|
|
| 2 |
import spaces
|
| 3 |
import torch
|
| 4 |
from diffusers import AutoencoderKL, TCDScheduler
|
| 5 |
-
|
| 6 |
-
from gradio_imageslider import ImageSlider
|
| 7 |
-
from huggingface_hub import hf_hub_download
|
| 8 |
-
|
| 9 |
from controlnet_union import ControlNetModel_Union
|
| 10 |
from pipeline_fill_sd_xl import StableDiffusionXLFillPipeline
|
|
|
|
|
|
|
| 11 |
|
| 12 |
from PIL import Image, ImageDraw
|
| 13 |
import numpy as np
|
| 14 |
|
| 15 |
-
|
|
|
|
|
|
|
| 16 |
"xinsir/controlnet-union-sdxl-1.0",
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
config = ControlNetModel_Union.load_config(config_file)
|
| 21 |
-
controlnet_model = ControlNetModel_Union.from_config(config)
|
| 22 |
-
model_file = hf_hub_download(
|
| 23 |
-
"xinsir/controlnet-union-sdxl-1.0",
|
| 24 |
-
filename="diffusion_pytorch_model_promax.safetensors",
|
| 25 |
-
)
|
| 26 |
-
|
| 27 |
-
sstate_dict = load_state_dict(model_file)
|
| 28 |
-
model, _, _, _, _ = ControlNetModel_Union._load_pretrained_model(
|
| 29 |
-
controlnet_model, sstate_dict, model_file, "xinsir/controlnet-union-sdxl-1.0"
|
| 30 |
-
)
|
| 31 |
-
model.to(device="cuda", dtype=torch.float16)
|
| 32 |
-
#----------------------
|
| 33 |
|
| 34 |
vae = AutoencoderKL.from_pretrained(
|
| 35 |
-
"madebyollin/sdxl-vae-fp16-fix",
|
|
|
|
| 36 |
).to("cuda")
|
| 37 |
|
| 38 |
pipe = StableDiffusionXLFillPipeline.from_pretrained(
|
| 39 |
"SG161222/RealVisXL_V5.0_Lightning",
|
| 40 |
torch_dtype=torch.float16,
|
| 41 |
vae=vae,
|
| 42 |
-
controlnet=
|
| 43 |
variant="fp16",
|
| 44 |
).to("cuda")
|
| 45 |
-
|
| 46 |
pipe.scheduler = TCDScheduler.from_config(pipe.scheduler.config)
|
| 47 |
|
|
|
|
| 48 |
def can_expand(source_width, source_height, target_width, target_height, alignment):
|
| 49 |
-
"""Checks if the image can be expanded based on the alignment."""
|
| 50 |
if alignment in ("Left", "Right") and source_width >= target_width:
|
| 51 |
return False
|
| 52 |
if alignment in ("Top", "Bottom") and source_height >= target_height:
|
| 53 |
return False
|
| 54 |
return True
|
| 55 |
|
| 56 |
-
def prepare_image_and_mask(image, width, height, overlap_percentage, resize_option, custom_resize_percentage, alignment, overlap_left, overlap_right, overlap_top, overlap_bottom):
|
| 57 |
-
target_size = (width, height)
|
| 58 |
-
|
| 59 |
-
# Calculate the scaling factor to fit the image within the target size
|
| 60 |
-
scale_factor = min(target_size[0] / image.width, target_size[1] / image.height)
|
| 61 |
-
new_width = int(image.width * scale_factor)
|
| 62 |
-
new_height = int(image.height * scale_factor)
|
| 63 |
-
|
| 64 |
-
# Resize the source image to fit within target size
|
| 65 |
-
source = image.resize((new_width, new_height), Image.LANCZOS)
|
| 66 |
-
|
| 67 |
-
# Apply resize option using percentages
|
| 68 |
-
if resize_option == "Full":
|
| 69 |
-
resize_percentage = 100
|
| 70 |
-
elif resize_option == "50%":
|
| 71 |
-
resize_percentage = 50
|
| 72 |
-
elif resize_option == "33%":
|
| 73 |
-
resize_percentage = 33
|
| 74 |
-
elif resize_option == "25%":
|
| 75 |
-
resize_percentage = 25
|
| 76 |
-
else: # Custom
|
| 77 |
-
resize_percentage = custom_resize_percentage
|
| 78 |
-
|
| 79 |
-
# Calculate new dimensions based on percentage
|
| 80 |
-
resize_factor = resize_percentage / 100
|
| 81 |
-
new_width = int(source.width * resize_factor)
|
| 82 |
-
new_height = int(source.height * resize_factor)
|
| 83 |
-
|
| 84 |
-
# Ensure minimum size of 64 pixels
|
| 85 |
-
new_width = max(new_width, 64)
|
| 86 |
-
new_height = max(new_height, 64)
|
| 87 |
-
|
| 88 |
-
# Resize the image
|
| 89 |
-
source = source.resize((new_width, new_height), Image.LANCZOS)
|
| 90 |
-
|
| 91 |
-
# Calculate the overlap in pixels based on the percentage
|
| 92 |
-
overlap_x = int(new_width * (overlap_percentage / 100))
|
| 93 |
-
overlap_y = int(new_height * (overlap_percentage / 100))
|
| 94 |
-
|
| 95 |
-
# Ensure minimum overlap of 1 pixel
|
| 96 |
-
overlap_x = max(overlap_x, 1)
|
| 97 |
-
overlap_y = max(overlap_y, 1)
|
| 98 |
-
|
| 99 |
-
# Calculate margins based on alignment
|
| 100 |
-
if alignment == "Middle":
|
| 101 |
-
margin_x = (target_size[0] - new_width) // 2
|
| 102 |
-
margin_y = (target_size[1] - new_height) // 2
|
| 103 |
-
elif alignment == "Left":
|
| 104 |
-
margin_x = 0
|
| 105 |
-
margin_y = (target_size[1] - new_height) // 2
|
| 106 |
-
elif alignment == "Right":
|
| 107 |
-
margin_x = target_size[0] - new_width
|
| 108 |
-
margin_y = (target_size[1] - new_height) // 2
|
| 109 |
-
elif alignment == "Top":
|
| 110 |
-
margin_x = (target_size[0] - new_width) // 2
|
| 111 |
-
margin_y = 0
|
| 112 |
-
elif alignment == "Bottom":
|
| 113 |
-
margin_x = (target_size[0] - new_width) // 2
|
| 114 |
-
margin_y = target_size[1] - new_height
|
| 115 |
-
|
| 116 |
-
# Adjust margins to eliminate gaps
|
| 117 |
-
margin_x = max(0, min(margin_x, target_size[0] - new_width))
|
| 118 |
-
margin_y = max(0, min(margin_y, target_size[1] - new_height))
|
| 119 |
-
|
| 120 |
-
# Create a new background image and paste the resized source image
|
| 121 |
-
background = Image.new('RGB', target_size, (255, 255, 255))
|
| 122 |
-
background.paste(source, (margin_x, margin_y))
|
| 123 |
-
|
| 124 |
-
# Create the mask
|
| 125 |
-
mask = Image.new('L', target_size, 255)
|
| 126 |
-
mask_draw = ImageDraw.Draw(mask)
|
| 127 |
-
|
| 128 |
-
# Calculate overlap areas
|
| 129 |
-
white_gaps_patch = 2
|
| 130 |
-
|
| 131 |
-
left_overlap = margin_x + overlap_x if overlap_left else margin_x + white_gaps_patch
|
| 132 |
-
right_overlap = margin_x + new_width - overlap_x if overlap_right else margin_x + new_width - white_gaps_patch
|
| 133 |
-
top_overlap = margin_y + overlap_y if overlap_top else margin_y + white_gaps_patch
|
| 134 |
-
bottom_overlap = margin_y + new_height - overlap_y if overlap_bottom else margin_y + new_height - white_gaps_patch
|
| 135 |
-
|
| 136 |
-
if alignment == "Left":
|
| 137 |
-
left_overlap = margin_x + overlap_x if overlap_left else margin_x
|
| 138 |
-
elif alignment == "Right":
|
| 139 |
-
right_overlap = margin_x + new_width - overlap_x if overlap_right else margin_x + new_width
|
| 140 |
-
elif alignment == "Top":
|
| 141 |
-
top_overlap = margin_y + overlap_y if overlap_top else margin_y
|
| 142 |
-
elif alignment == "Bottom":
|
| 143 |
-
bottom_overlap = margin_y + new_height - overlap_y if overlap_bottom else margin_y + new_height
|
| 144 |
-
|
| 145 |
-
|
| 146 |
-
# Draw the mask
|
| 147 |
-
mask_draw.rectangle([
|
| 148 |
-
(left_overlap, top_overlap),
|
| 149 |
-
(right_overlap, bottom_overlap)
|
| 150 |
-
], fill=0)
|
| 151 |
-
|
| 152 |
-
return background, mask
|
| 153 |
-
|
| 154 |
-
def preview_image_and_mask(image, width, height, overlap_percentage, resize_option, custom_resize_percentage, alignment, overlap_left, overlap_right, overlap_top, overlap_bottom):
|
| 155 |
-
background, mask = prepare_image_and_mask(image, width, height, overlap_percentage, resize_option, custom_resize_percentage, alignment, overlap_left, overlap_right, overlap_top, overlap_bottom)
|
| 156 |
-
|
| 157 |
-
# Create a preview image showing the mask
|
| 158 |
-
preview = background.copy().convert('RGBA')
|
| 159 |
-
|
| 160 |
-
# Create a semi-transparent red overlay
|
| 161 |
-
red_overlay = Image.new('RGBA', background.size, (255, 0, 0, 64)) # Reduced alpha to 64 (25% opacity)
|
| 162 |
-
|
| 163 |
-
# Convert black pixels in the mask to semi-transparent red
|
| 164 |
-
red_mask = Image.new('RGBA', background.size, (0, 0, 0, 0))
|
| 165 |
-
red_mask.paste(red_overlay, (0, 0), mask)
|
| 166 |
-
|
| 167 |
-
# Overlay the red mask on the background
|
| 168 |
-
preview = Image.alpha_composite(preview, red_mask)
|
| 169 |
-
|
| 170 |
-
return preview
|
| 171 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 172 |
@spaces.GPU(duration=24)
|
| 173 |
-
def infer(image, width, height, overlap_percentage, num_inference_steps,
|
| 174 |
-
|
| 175 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 176 |
if not can_expand(background.width, background.height, width, height, alignment):
|
| 177 |
alignment = "Middle"
|
| 178 |
|
| 179 |
-
|
| 180 |
-
|
| 181 |
|
| 182 |
final_prompt = f"{prompt_input} , high quality, 4k"
|
| 183 |
-
|
| 184 |
-
|
| 185 |
-
|
| 186 |
-
|
| 187 |
-
|
| 188 |
-
|
| 189 |
-
|
| 190 |
-
|
| 191 |
-
|
| 192 |
-
prompt_embeds=prompt_embeds,
|
| 193 |
-
negative_prompt_embeds=negative_prompt_embeds,
|
| 194 |
-
pooled_prompt_embeds=pooled_prompt_embeds,
|
| 195 |
-
negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
|
| 196 |
-
image=cnet_image,
|
| 197 |
num_inference_steps=num_inference_steps
|
| 198 |
-
)
|
| 199 |
-
|
|
|
|
|
|
|
| 200 |
|
| 201 |
-
|
| 202 |
-
|
|
|
|
|
|
|
|
|
|
| 203 |
|
| 204 |
-
yield background, cnet_image
|
| 205 |
|
| 206 |
def clear_result():
|
| 207 |
-
"""Clears the result ImageSlider."""
|
| 208 |
return gr.update(value=None)
|
| 209 |
|
| 210 |
-
def preload_presets(
|
| 211 |
-
|
| 212 |
-
if
|
| 213 |
-
|
| 214 |
-
|
| 215 |
-
return changed_width, changed_height, gr.update()
|
| 216 |
-
elif target_ratio == "16:9":
|
| 217 |
-
changed_width = 1280
|
| 218 |
-
changed_height = 720
|
| 219 |
-
return changed_width, changed_height, gr.update()
|
| 220 |
-
elif target_ratio == "1:1":
|
| 221 |
-
changed_width = 1024
|
| 222 |
-
changed_height = 1024
|
| 223 |
-
return changed_width, changed_height, gr.update()
|
| 224 |
-
elif target_ratio == "Custom":
|
| 225 |
-
return ui_width, ui_height, gr.update(open=True)
|
| 226 |
-
|
| 227 |
-
def select_the_right_preset(user_width, user_height):
|
| 228 |
-
if user_width == 720 and user_height == 1280:
|
| 229 |
-
return "9:16"
|
| 230 |
-
elif user_width == 1280 and user_height == 720:
|
| 231 |
-
return "16:9"
|
| 232 |
-
elif user_width == 1024 and user_height == 1024:
|
| 233 |
-
return "1:1"
|
| 234 |
-
else:
|
| 235 |
-
return "Custom"
|
| 236 |
-
|
| 237 |
-
def toggle_custom_resize_slider(resize_option):
|
| 238 |
-
return gr.update(visible=(resize_option == "Custom"))
|
| 239 |
-
|
| 240 |
-
def update_history(new_image, history):
|
| 241 |
-
"""Updates the history gallery with the new image."""
|
| 242 |
-
if history is None:
|
| 243 |
-
history = []
|
| 244 |
-
history.insert(0, new_image)
|
| 245 |
-
return history
|
| 246 |
|
| 247 |
-
|
| 248 |
-
|
| 249 |
-
|
| 250 |
-
|
| 251 |
-
|
| 252 |
-
footer { visibility: hidden; }
|
| 253 |
-
"""
|
| 254 |
|
| 255 |
-
|
| 256 |
-
""
|
| 257 |
|
| 258 |
-
|
| 259 |
-
|
| 260 |
-
|
| 261 |
-
|
| 262 |
-
with gr.Row():
|
| 263 |
-
with gr.Column():
|
| 264 |
-
input_image = gr.Image(
|
| 265 |
-
type="pil",
|
| 266 |
-
label="Input Image"
|
| 267 |
-
)
|
| 268 |
-
|
| 269 |
-
with gr.Row():
|
| 270 |
-
with gr.Column(scale=2):
|
| 271 |
-
prompt_input = gr.Textbox(label="Prompt (Optional)")
|
| 272 |
-
with gr.Column(scale=1):
|
| 273 |
-
run_button = gr.Button("Generate")
|
| 274 |
-
|
| 275 |
-
with gr.Row():
|
| 276 |
-
target_ratio = gr.Radio(
|
| 277 |
-
label="Expected Ratio",
|
| 278 |
-
choices=["9:16", "16:9", "1:1", "Custom"],
|
| 279 |
-
value="9:16",
|
| 280 |
-
scale=2
|
| 281 |
-
)
|
| 282 |
-
|
| 283 |
-
alignment_dropdown = gr.Dropdown(
|
| 284 |
-
choices=["Middle", "Left", "Right", "Top", "Bottom"],
|
| 285 |
-
value="Middle",
|
| 286 |
-
label="Alignment"
|
| 287 |
-
)
|
| 288 |
-
|
| 289 |
-
with gr.Accordion(label="Advanced settings", open=False) as settings_panel:
|
| 290 |
-
with gr.Column():
|
| 291 |
-
with gr.Row():
|
| 292 |
-
width_slider = gr.Slider(
|
| 293 |
-
label="Target Width",
|
| 294 |
-
minimum=720,
|
| 295 |
-
maximum=1536,
|
| 296 |
-
step=8,
|
| 297 |
-
value=720, # Set a default value
|
| 298 |
-
)
|
| 299 |
-
height_slider = gr.Slider(
|
| 300 |
-
label="Target Height",
|
| 301 |
-
minimum=720,
|
| 302 |
-
maximum=1536,
|
| 303 |
-
step=8,
|
| 304 |
-
value=1280, # Set a default value
|
| 305 |
-
)
|
| 306 |
-
|
| 307 |
-
num_inference_steps = gr.Slider(label="Steps", minimum=4, maximum=12, step=1, value=8)
|
| 308 |
-
with gr.Group():
|
| 309 |
-
overlap_percentage = gr.Slider(
|
| 310 |
-
label="Mask overlap (%)",
|
| 311 |
-
minimum=1,
|
| 312 |
-
maximum=50,
|
| 313 |
-
value=10,
|
| 314 |
-
step=1
|
| 315 |
-
)
|
| 316 |
-
with gr.Row():
|
| 317 |
-
overlap_top = gr.Checkbox(label="Overlap Top", value=True)
|
| 318 |
-
overlap_right = gr.Checkbox(label="Overlap Right", value=True)
|
| 319 |
-
with gr.Row():
|
| 320 |
-
overlap_left = gr.Checkbox(label="Overlap Left", value=True)
|
| 321 |
-
overlap_bottom = gr.Checkbox(label="Overlap Bottom", value=True)
|
| 322 |
-
with gr.Row():
|
| 323 |
-
resize_option = gr.Radio(
|
| 324 |
-
label="Resize input image",
|
| 325 |
-
choices=["Full", "50%", "33%", "25%", "Custom"],
|
| 326 |
-
value="Full"
|
| 327 |
-
)
|
| 328 |
-
custom_resize_percentage = gr.Slider(
|
| 329 |
-
label="Custom resize (%)",
|
| 330 |
-
minimum=1,
|
| 331 |
-
maximum=100,
|
| 332 |
-
step=1,
|
| 333 |
-
value=50,
|
| 334 |
-
visible=False
|
| 335 |
-
)
|
| 336 |
-
|
| 337 |
-
with gr.Column():
|
| 338 |
-
preview_button = gr.Button("Preview alignment and mask")
|
| 339 |
-
|
| 340 |
-
|
| 341 |
-
gr.Examples(
|
| 342 |
-
examples=[
|
| 343 |
-
["./examples/example_1.webp", 1280, 720, "Middle"],
|
| 344 |
-
["./examples/example_2.jpg", 1440, 810, "Left"],
|
| 345 |
-
["./examples/example_3.jpg", 1024, 1024, "Top"],
|
| 346 |
-
["./examples/example_3.jpg", 1024, 1024, "Bottom"],
|
| 347 |
-
],
|
| 348 |
-
inputs=[input_image, width_slider, height_slider, alignment_dropdown],
|
| 349 |
-
)
|
| 350 |
-
|
| 351 |
-
|
| 352 |
-
|
| 353 |
-
with gr.Column():
|
| 354 |
-
result = ImageSlider(label="Generated Image", interactive=False, type="pil", slider_color="pink")
|
| 355 |
-
use_as_input_button = gr.Button("Use as Input Image", visible=False)
|
| 356 |
-
|
| 357 |
-
history_gallery = gr.Gallery(label="History", columns=6, object_fit="contain", interactive=False)
|
| 358 |
-
preview_image = gr.Image(label="Preview")
|
| 359 |
-
|
| 360 |
-
|
| 361 |
-
|
| 362 |
-
def use_output_as_input(output_image):
|
| 363 |
-
"""Sets the generated output as the new input image."""
|
| 364 |
-
return gr.update(value=output_image[1])
|
| 365 |
-
|
| 366 |
-
use_as_input_button.click(
|
| 367 |
-
fn=use_output_as_input,
|
| 368 |
-
inputs=[result],
|
| 369 |
-
outputs=[input_image]
|
| 370 |
-
)
|
| 371 |
-
|
| 372 |
-
target_ratio.change(
|
| 373 |
-
fn=preload_presets,
|
| 374 |
-
inputs=[target_ratio, width_slider, height_slider],
|
| 375 |
-
outputs=[width_slider, height_slider, settings_panel],
|
| 376 |
-
queue=False
|
| 377 |
-
)
|
| 378 |
-
|
| 379 |
-
width_slider.change(
|
| 380 |
-
fn=select_the_right_preset,
|
| 381 |
-
inputs=[width_slider, height_slider],
|
| 382 |
-
outputs=[target_ratio],
|
| 383 |
-
queue=False
|
| 384 |
-
)
|
| 385 |
-
|
| 386 |
-
height_slider.change(
|
| 387 |
-
fn=select_the_right_preset,
|
| 388 |
-
inputs=[width_slider, height_slider],
|
| 389 |
-
outputs=[target_ratio],
|
| 390 |
-
queue=False
|
| 391 |
-
)
|
| 392 |
|
| 393 |
-
|
| 394 |
-
|
| 395 |
-
inputs=[resize_option],
|
| 396 |
-
outputs=[custom_resize_percentage],
|
| 397 |
-
queue=False
|
| 398 |
-
)
|
| 399 |
-
|
| 400 |
-
run_button.click( # Clear the result
|
| 401 |
-
fn=clear_result,
|
| 402 |
-
inputs=None,
|
| 403 |
-
outputs=result,
|
| 404 |
-
).then( # Generate the new image
|
| 405 |
-
fn=infer,
|
| 406 |
-
inputs=[input_image, width_slider, height_slider, overlap_percentage, num_inference_steps,
|
| 407 |
-
resize_option, custom_resize_percentage, prompt_input, alignment_dropdown,
|
| 408 |
-
overlap_left, overlap_right, overlap_top, overlap_bottom],
|
| 409 |
-
outputs=result,
|
| 410 |
-
).then( # Update the history gallery
|
| 411 |
-
fn=lambda x, history: update_history(x[1], history),
|
| 412 |
-
inputs=[result, history_gallery],
|
| 413 |
-
outputs=history_gallery,
|
| 414 |
-
).then( # Show the "Use as Input Image" button
|
| 415 |
-
fn=lambda: gr.update(visible=True),
|
| 416 |
-
inputs=None,
|
| 417 |
-
outputs=use_as_input_button,
|
| 418 |
-
)
|
| 419 |
|
| 420 |
-
|
| 421 |
-
|
| 422 |
-
|
| 423 |
-
|
| 424 |
-
|
| 425 |
-
|
| 426 |
-
|
| 427 |
-
|
| 428 |
-
|
| 429 |
-
|
| 430 |
-
|
| 431 |
-
|
| 432 |
-
|
| 433 |
-
|
| 434 |
-
|
| 435 |
-
|
| 436 |
-
|
| 437 |
-
|
| 438 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 439 |
|
| 440 |
-
preview_button.click(
|
| 441 |
-
fn=preview_image_and_mask,
|
| 442 |
-
inputs=[input_image, width_slider, height_slider, overlap_percentage, resize_option, custom_resize_percentage, alignment_dropdown,
|
| 443 |
-
overlap_left, overlap_right, overlap_top, overlap_bottom],
|
| 444 |
-
outputs=preview_image,
|
| 445 |
-
queue=False
|
| 446 |
-
)
|
| 447 |
|
| 448 |
demo.queue(max_size=20).launch(share=False, ssr_mode=False, show_error=True)
|
|
|
|
| 2 |
import spaces
|
| 3 |
import torch
|
| 4 |
from diffusers import AutoencoderKL, TCDScheduler
|
| 5 |
+
# (Assume ControlNet manual load or from_pretrained is already working)
|
|
|
|
|
|
|
|
|
|
| 6 |
from controlnet_union import ControlNetModel_Union
|
| 7 |
from pipeline_fill_sd_xl import StableDiffusionXLFillPipeline
|
| 8 |
+
from gradio_imageslider import ImageSlider
|
| 9 |
+
from huggingface_hub import hf_hub_download
|
| 10 |
|
| 11 |
from PIL import Image, ImageDraw
|
| 12 |
import numpy as np
|
| 13 |
|
| 14 |
+
# --- Load ControlNet and SDXL Fill Pipeline ---
|
| 15 |
+
# (Either manual download or via from_pretrained)
|
| 16 |
+
controlnet_model = ControlNetModel_Union.from_pretrained(
|
| 17 |
"xinsir/controlnet-union-sdxl-1.0",
|
| 18 |
+
torch_dtype=torch.float16,
|
| 19 |
+
variant="fp16"
|
| 20 |
+
).to("cuda")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
|
| 22 |
vae = AutoencoderKL.from_pretrained(
|
| 23 |
+
"madebyollin/sdxl-vae-fp16-fix",
|
| 24 |
+
torch_dtype=torch.float16
|
| 25 |
).to("cuda")
|
| 26 |
|
| 27 |
pipe = StableDiffusionXLFillPipeline.from_pretrained(
|
| 28 |
"SG161222/RealVisXL_V5.0_Lightning",
|
| 29 |
torch_dtype=torch.float16,
|
| 30 |
vae=vae,
|
| 31 |
+
controlnet=controlnet_model,
|
| 32 |
variant="fp16",
|
| 33 |
).to("cuda")
|
|
|
|
| 34 |
pipe.scheduler = TCDScheduler.from_config(pipe.scheduler.config)
|
| 35 |
|
| 36 |
+
# --- Utility functions ---
|
| 37 |
def can_expand(source_width, source_height, target_width, target_height, alignment):
|
|
|
|
| 38 |
if alignment in ("Left", "Right") and source_width >= target_width:
|
| 39 |
return False
|
| 40 |
if alignment in ("Top", "Bottom") and source_height >= target_height:
|
| 41 |
return False
|
| 42 |
return True
|
| 43 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 44 |
|
| 45 |
+
def prepare_image_and_mask(image, width, height, overlap_percentage,
|
| 46 |
+
resize_option, custom_resize_percentage,
|
| 47 |
+
alignment, overlap_left, overlap_right,
|
| 48 |
+
overlap_top, overlap_bottom):
|
| 49 |
+
target = (width, height)
|
| 50 |
+
scale = min(target[0] / image.width, target[1] / image.height)
|
| 51 |
+
w, h = int(image.width * scale), int(image.height * scale)
|
| 52 |
+
src = image.resize((w, h), Image.LANCZOS)
|
| 53 |
+
|
| 54 |
+
# Resize percentage
|
| 55 |
+
if resize_option == "Full": pct = 100
|
| 56 |
+
elif resize_option == "50%": pct = 50
|
| 57 |
+
elif resize_option == "33%": pct = 33
|
| 58 |
+
elif resize_option == "25%": pct = 25
|
| 59 |
+
else: pct = custom_resize_percentage
|
| 60 |
+
|
| 61 |
+
rw, rh = max(int(src.width * pct / 100), 64), max(int(src.height * pct / 100), 64)
|
| 62 |
+
src = src.resize((rw, rh), Image.LANCZOS)
|
| 63 |
+
|
| 64 |
+
ox = max(int(rw * overlap_percentage / 100), 1)
|
| 65 |
+
oy = max(int(rh * overlap_percentage / 100), 1)
|
| 66 |
+
|
| 67 |
+
# Margins
|
| 68 |
+
if alignment == "Middle": mx, my = (width - rw)//2, (height - rh)//2
|
| 69 |
+
elif alignment == "Left": mx, my = 0, (height - rh)//2
|
| 70 |
+
elif alignment == "Right": mx, my = width - rw, (height - rh)//2
|
| 71 |
+
elif alignment == "Top": mx, my = (width - rw)//2, 0
|
| 72 |
+
else: mx, my = (width - rw)//2, height - rh
|
| 73 |
+
|
| 74 |
+
mx, my = max(0, min(mx, width - rw)), max(0, min(my, height - rh))
|
| 75 |
+
|
| 76 |
+
bg = Image.new("RGB", target, (255,255,255))
|
| 77 |
+
bg.paste(src, (mx, my))
|
| 78 |
+
|
| 79 |
+
mask = Image.new("L", target, 255)
|
| 80 |
+
d = ImageDraw.Draw(mask)
|
| 81 |
+
|
| 82 |
+
lx = mx + (ox if overlap_left else 2)
|
| 83 |
+
rx = mx + rw - (ox if overlap_right else 2)
|
| 84 |
+
ty = my + (oy if overlap_top else 2)
|
| 85 |
+
by = my + rh - (oy if overlap_bottom else 2)
|
| 86 |
+
|
| 87 |
+
# Edge adjustments
|
| 88 |
+
if alignment == "Left": lx = mx + (ox if overlap_left else 0)
|
| 89 |
+
if alignment == "Right": rx = mx + rw - (ox if overlap_right else 0)
|
| 90 |
+
if alignment == "Top": ty = my + (oy if overlap_top else 0)
|
| 91 |
+
if alignment == "Bottom": by = my + rh - (oy if overlap_bottom else 0)
|
| 92 |
+
|
| 93 |
+
d.rectangle([(lx, ty), (rx, by)], fill=0)
|
| 94 |
+
return bg, mask
|
| 95 |
+
|
| 96 |
+
|
| 97 |
+
def preview_image_and_mask(*args):
|
| 98 |
+
bg, mask = prepare_image_and_mask(*args)
|
| 99 |
+
vis = bg.copy().convert("RGBA")
|
| 100 |
+
red = Image.new("RGBA", bg.size, (255,0,0,64))
|
| 101 |
+
overlay = Image.new("RGBA", bg.size, (0,0,0,0))
|
| 102 |
+
overlay.paste(red, (0,0), mask)
|
| 103 |
+
return Image.alpha_composite(vis, overlay)
|
| 104 |
+
|
| 105 |
+
# --- Fixed infer: return list for slider ---
|
| 106 |
@spaces.GPU(duration=24)
|
| 107 |
+
def infer(image, width, height, overlap_percentage, num_inference_steps,
|
| 108 |
+
resize_option, custom_resize_percentage, prompt_input,
|
| 109 |
+
alignment, overlap_left, overlap_right, overlap_top, overlap_bottom):
|
| 110 |
+
background, mask = prepare_image_and_mask(
|
| 111 |
+
image, width, height, overlap_percentage,
|
| 112 |
+
resize_option, custom_resize_percentage,
|
| 113 |
+
alignment, overlap_left, overlap_right,
|
| 114 |
+
overlap_top, overlap_bottom
|
| 115 |
+
)
|
| 116 |
if not can_expand(background.width, background.height, width, height, alignment):
|
| 117 |
alignment = "Middle"
|
| 118 |
|
| 119 |
+
hole = background.copy()
|
| 120 |
+
hole.paste(0, (0,0), mask)
|
| 121 |
|
| 122 |
final_prompt = f"{prompt_input} , high quality, 4k"
|
| 123 |
+
embeds = pipe.encode_prompt(final_prompt, "cuda", True)
|
| 124 |
+
|
| 125 |
+
# Run pipeline and grab last frame
|
| 126 |
+
gen = pipe(
|
| 127 |
+
prompt_embeds=embeds[0],
|
| 128 |
+
negative_prompt_embeds=embeds[1],
|
| 129 |
+
pooled_prompt_embeds=embeds[2],
|
| 130 |
+
negative_pooled_prompt_embeds=embeds[3],
|
| 131 |
+
image=hole,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 132 |
num_inference_steps=num_inference_steps
|
| 133 |
+
)
|
| 134 |
+
last = None
|
| 135 |
+
for img in gen:
|
| 136 |
+
last = img
|
| 137 |
|
| 138 |
+
out = last.convert("RGBA")
|
| 139 |
+
hole.paste(out, (0,0), mask)
|
| 140 |
+
|
| 141 |
+
# Return a list: [input_hole_image, final_output]
|
| 142 |
+
return [background, hole]
|
| 143 |
|
|
|
|
| 144 |
|
| 145 |
def clear_result():
|
|
|
|
| 146 |
return gr.update(value=None)
|
| 147 |
|
| 148 |
+
def preload_presets(ratio, w, h):
|
| 149 |
+
if ratio == "9:16": return 720, 1280, gr.update()
|
| 150 |
+
if ratio == "16:9": return 1280, 720, gr.update()
|
| 151 |
+
if ratio == "1:1": return 1024, 1024, gr.update()
|
| 152 |
+
return w, h, gr.update(open=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 153 |
|
| 154 |
+
def select_the_right_preset(w, h):
|
| 155 |
+
if (w,h) == (720,1280): return "9:16"
|
| 156 |
+
if (w,h) == (1280,720): return "16:9"
|
| 157 |
+
if (w,h) == (1024,1024): return "1:1"
|
| 158 |
+
return "Custom"
|
|
|
|
|
|
|
| 159 |
|
| 160 |
+
def toggle_custom_resize_slider(opt):
|
| 161 |
+
return gr.update(visible=(opt=="Custom"))
|
| 162 |
|
| 163 |
+
def update_history(img, history):
|
| 164 |
+
history = history or []
|
| 165 |
+
history.insert(0, img)
|
| 166 |
+
return history
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 167 |
|
| 168 |
+
css = ".gradio-container { width: 1200px !important; }"
|
| 169 |
+
title = "<h1 align='center'>Diffusers Image Outpaint Lightning</h1>"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 170 |
|
| 171 |
+
with gr.Blocks(css=css) as demo:
|
| 172 |
+
gr.HTML(title)
|
| 173 |
+
with gr.Row():
|
| 174 |
+
with gr.Column():
|
| 175 |
+
input_image = gr.Image(type="pil", label="Input Image")
|
| 176 |
+
prompt_input = gr.Textbox(label="Prompt (Optional)")
|
| 177 |
+
run_button = gr.Button("Generate")
|
| 178 |
+
|
| 179 |
+
target_ratio = gr.Radio(["9:16","16:9","1:1","Custom"], value="9:16", label="Expected Ratio")
|
| 180 |
+
alignment_dropdown = gr.Dropdown(["Middle","Left","Right","Top","Bottom"], value="Middle", label="Alignment")
|
| 181 |
+
|
| 182 |
+
with gr.Accordion("Advanced settings", open=False) as adv:
|
| 183 |
+
width_slider = gr.Slider(720,1536,step=8, value=720, label="Target Width")
|
| 184 |
+
height_slider = gr.Slider(720,1536,step=8, value=1280, label="Target Height")
|
| 185 |
+
num_steps = gr.Slider(4,12,step=1, value=8, label="Steps")
|
| 186 |
+
overlap_pct = gr.Slider(1,50,step=1, value=10, label="Mask overlap (%)")
|
| 187 |
+
overlap_top = gr.Checkbox(label="Overlap Top", value=True)
|
| 188 |
+
overlap_right = gr.Checkbox(label="Overlap Right", value=True)
|
| 189 |
+
overlap_left = gr.Checkbox(label="Overlap Left", value=True)
|
| 190 |
+
overlap_bottom= gr.Checkbox(label="Overlap Bottom", value=True)
|
| 191 |
+
resize_opt = gr.Radio(["Full","50%","33%","25%","Custom"], value="Full", label="Resize input image")
|
| 192 |
+
custom_resize = gr.Slider(1,100,step=1, value=50, visible=False, label="Custom resize (%)")
|
| 193 |
+
preview_btn = gr.Button("Preview alignment and mask")
|
| 194 |
+
|
| 195 |
+
gr.Examples(
|
| 196 |
+
examples=[
|
| 197 |
+
["./examples/example_1.webp",1280,720,"Middle"],
|
| 198 |
+
["./examples/example_2.jpg",1440,810,"Left"],
|
| 199 |
+
["./examples/example_3.jpg",1024,1024,"Top"],
|
| 200 |
+
["./examples/example_3.jpg",1024,1024,"Bottom"]
|
| 201 |
+
],
|
| 202 |
+
inputs=[input_image,width_slider,height_slider,alignment_dropdown]
|
| 203 |
+
)
|
| 204 |
+
|
| 205 |
+
with gr.Column():
|
| 206 |
+
result = ImageSlider(label="Comparison", interactive=False, type="pil", slider_color="pink")
|
| 207 |
+
history_gallery = gr.Gallery(label="History", columns=6, object_fit="contain")
|
| 208 |
+
preview_image = gr.Image(label="Preview")
|
| 209 |
+
|
| 210 |
+
# Callbacks
|
| 211 |
+
run_button.click(clear_result, None, result)
|
| 212 |
+
run_button.click(
|
| 213 |
+
infer,
|
| 214 |
+
inputs=[ input_image, width_slider, height_slider, overlap_pct, num_steps,
|
| 215 |
+
resize_opt, custom_resize, prompt_input, alignment_dropdown,
|
| 216 |
+
overlap_left, overlap_right, overlap_top, overlap_bottom],
|
| 217 |
+
outputs=result
|
| 218 |
+
).then(update_history, inputs=[result, history_gallery], outputs=history_gallery)
|
| 219 |
+
|
| 220 |
+
target_ratio.change(preload_presets, [target_ratio, width_slider, height_slider], [width_slider, height_slider, adv])
|
| 221 |
+
width_slider.change(select_the_right_preset, [width_slider, height_slider], target_ratio)
|
| 222 |
+
height_slider.change(select_the_right_preset, [width_slider, height_slider], target_ratio)
|
| 223 |
+
resize_opt.change(toggle_custom_resize_slider, resize_opt, custom_resize)
|
| 224 |
+
preview_btn.click(preview_image_and_mask,
|
| 225 |
+
[input_image, width_slider, height_slider, overlap_pct, resize_opt, custom_resize, alignment_dropdown,
|
| 226 |
+
overlap_left, overlap_right, overlap_top, overlap_bottom],
|
| 227 |
+
preview_image)
|
| 228 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 229 |
|
| 230 |
demo.queue(max_size=20).launch(share=False, ssr_mode=False, show_error=True)
|