Spaces:
Running
on
Zero
Running
on
Zero
File size: 16,086 Bytes
d65c9b3 657e152 d65c9b3 657e152 d65c9b3 657e152 d65c9b3 abc734b d65c9b3 ff54977 d65c9b3 ff54977 d65c9b3 ff54977 d65c9b3 ff54977 d65c9b3 ff54977 abc734b d65c9b3 ff54977 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 |
from __future__ import annotations
import gradio as gr
import spaces
from PIL import Image
import torch
from src.eunms import Model_Type, Scheduler_Type, Gradient_Averaging_Type, Epsilon_Update_Type
from src.enums_utils import model_type_to_size, get_pipes
from src.config import RunConfig
from main import run as run_model
DESCRIPTION = '''# ReNoise: Real Image Inversion Through Iterative Noising
This is a demo for our ''ReNoise: Real Image Inversion Through Iterative Noising'' [paper](https://garibida.github.io/ReNoise-Inversion/). Code is available [here](https://github.com/garibida/ReNoise-Inversion)
Our ReNoise inversion technique can be applied to various diffusion models, including recent few-step ones such as SDXL-Turbo.
This demo preform real image editing using our ReNoise inversion. The input image is resize to size of 512x512, the optimal size of SDXL Turbo.
'''
device = 'cuda' if torch.cuda.is_available() else 'cpu'
model_type = Model_Type.SDXL_Turbo
scheduler_type = Scheduler_Type.EULER
image_size = model_type_to_size(Model_Type.SDXL_Turbo)
pipe_inversion, pipe_inference = get_pipes(model_type, scheduler_type, device=device)
cache_size = 25
prev_configs = [None for i in range(cache_size)]
prev_inv_latents = [None for i in range(cache_size)]
prev_images = [None for i in range(cache_size)]
prev_noises = [None for i in range(cache_size)]
@spaces.GPU
def main_pipeline(
input_image: str,
src_prompt: str,
tgt_prompt: str,
edit_cfg: float,
number_of_renoising_iterations: int,
inersion_strength: float,
avg_gradients: bool,
first_step_range_start: int,
first_step_range_end: int,
rest_step_range_start: int,
rest_step_range_end: int,
lambda_ac: float,
lambda_kl: float,
noise_correction: bool):
global prev_configs, prev_inv_latents, prev_images, prev_noises
update_epsilon_type = Epsilon_Update_Type.OPTIMIZE if noise_correction else Epsilon_Update_Type.NONE
avg_gradients_type = Gradient_Averaging_Type.ON_END if avg_gradients else Gradient_Averaging_Type.NONE
first_step_range = (first_step_range_start, first_step_range_end)
rest_step_range = (rest_step_range_start, rest_step_range_end)
config = RunConfig(model_type = model_type,
num_inference_steps = 4,
num_inversion_steps = 4,
guidance_scale = 0.0,
max_num_aprox_steps_first_step = first_step_range_end+1,
num_aprox_steps = number_of_renoising_iterations,
inversion_max_step = inersion_strength,
gradient_averaging_type = avg_gradients_type,
gradient_averaging_first_step_range = first_step_range,
gradient_averaging_step_range = rest_step_range,
scheduler_type = scheduler_type,
num_reg_steps = 4,
num_ac_rolls = 5,
lambda_ac = lambda_ac,
lambda_kl = lambda_kl,
update_epsilon_type = update_epsilon_type,
do_reconstruction = True)
config.prompt = src_prompt
inv_latent = None
noise_list = None
for i in range(cache_size):
if prev_configs[i] is not None and prev_configs[i] == config and prev_images[i] == input_image:
print(f"Using cache for config #{i}")
inv_latent = prev_inv_latents[i]
noise_list = prev_noises[i]
prev_configs.pop(i)
prev_inv_latents.pop(i)
prev_images.pop(i)
prev_noises.pop(i)
break
original_image = Image.open(input_image).convert("RGB").resize(image_size)
res_image, inv_latent, noise, all_latents = run_model(original_image,
config,
latents=inv_latent,
pipe_inversion=pipe_inversion,
pipe_inference=pipe_inference,
edit_prompt=tgt_prompt,
noise=noise_list,
edit_cfg=edit_cfg)
prev_configs.append(config)
prev_inv_latents.append(inv_latent)
prev_images.append(input_image)
prev_noises.append(noise)
if len(prev_configs) > cache_size:
print("Popping cache")
prev_configs.pop(0)
prev_inv_latents.pop(0)
prev_images.pop(0)
prev_noises.pop(0)
return res_image
with gr.Blocks(css='style.css') as demo:
gr.Markdown(DESCRIPTION)
gr.HTML(
'''<a href="https://huggingface.co/spaces/garibida/ReNoise-Inversion?duplicate=true">
<img src="https://bit.ly/3gLdBN6" alt="Duplicate Space"></a>Duplicate the Space to run privately without waiting in queue''')
with gr.Row():
with gr.Column():
input_image = gr.Image(
label="Input image",
type="filepath",
height=image_size[0],
width=image_size[1]
)
src_prompt = gr.Text(
label='Source Prompt',
max_lines=1,
placeholder='A kitten is sitting in a basket on a branch',
)
tgt_prompt = gr.Text(
label='Target Prompt',
max_lines=1,
placeholder='A plush toy kitten is sitting in a basket on a branch',
)
with gr.Accordion("Advanced Options", open=False):
edit_cfg = gr.Slider(
label='Denoise Classifier-Free Guidence Scale',
minimum=1.0,
maximum=3.5,
value=1.0,
step=0.1
)
number_of_renoising_iterations = gr.Slider(
label='Number of ReNoise Iterations',
minimum=0,
maximum=20,
value=9,
step=1
)
inersion_strength = gr.Slider(
label='Inversion Strength',
info="Indicates how much to invert the reference image. The number of denoising steps depends on the amount of noise initially added. When strength is 1, the image will be inverted to complete noise and the denoising process will run for the full number of steps (4). When strength is 0.5, the image will be inverted to half noise and the denoising process will run for 2 steps.",
minimum=0.0,
maximum=1.0,
value=1.0,
step=0.25
)
avg_gradients = gr.Checkbox(
label="Preform Estimation Averaging",
info="IMPROVES RECONSTRUCTION. Averagin the estination over multiple ReNoise iterations can improve the quality of the reconstruction. The Next 4 sliders control the range of steps to average over. The first two sliders control the range of steps to average over for the first inversion step (t < 250). The last two sliders control the range of steps to average over for the rest of the inversion step (t > 250).",
value=True
)
first_step_range_start = gr.Slider(
label='First Estimation in Average (t < 250)',
minimum=0,
maximum=21,
value=0,
step=1
)
first_step_range_end = gr.Slider(
label='Last Estimation in Average (t < 250)',
minimum=0,
maximum=21,
value=5,
step=1
)
rest_step_range_start = gr.Slider(
label='First Estimation in Average (t > 250)',
minimum=0,
maximum=21,
value=8,
step=1
)
rest_step_range_end = gr.Slider(
label='Last Estimation in Average (t > 250)',
minimum=0,
maximum=21,
value=10,
step=1
)
num_reg_steps = 4
num_ac_rolls = 5
lambda_ac = gr.Slider(
label='Labmda AC',
info="IMPROVES EDITABILITY. The weight of the pair loss in the noise prediction regulariztion. This loss encourages the inversion to predict more editable noise. A higher value allows more significant changes to the image (higher editability), but may result in less faithful reconstructions.",
minimum=0.0,
maximum=50.0,
value=20.0,
step=1.0
)
lambda_kl = gr.Slider(
label='Labmda Patch KL',
info="IMPROVES EDITABILITY. This weight controls the strength of the patch-level KL divergence term in the noise prediction regularization. While it encourages editable noise like the Labmda AC, it often has a less detrimental effect on reconstruction fidelity.",
minimum=0.0,
maximum=0.4,
value=0.065,
step=0.005
)
noise_correction = gr.Checkbox(
label="Preform Noise Correction",
info="IMPROVES RECONSTRUCTION. Performs noise correction to improve the reconstruction of the image.",
value=True
)
run_button = gr.Button('Edit')
with gr.Column():
# result = gr.Gallery(label='Result')
result = gr.Image(
label="Result",
type="pil",
height=image_size[0],
width=image_size[1]
)
examples = [
[
"example_images/kitten.jpg", #input_image
"A kitten is sitting in a basket on a branch", #src_prompt
"a lego kitten is sitting in a basket on a branch", #tgt_prompt
1.0, #edit_cfg
9, #number_of_renoising_iterations
1.0, #inersion_strength
True, #avg_gradients
0, #first_step_range_start
5, #first_step_range_end
8, #rest_step_range_start
10, #rest_step_range_end
20.0, #lambda_ac
0.055, #lambda_kl
False #noise_correction
],
[
"example_images/kitten.jpg", #input_image
"A kitten is sitting in a basket on a branch", #src_prompt
"a brokkoli is sitting in a basket on a branch", #tgt_prompt
1.0, #edit_cfg
9, #number_of_renoising_iterations
1.0, #inersion_strength
True, #avg_gradients
0, #first_step_range_start
5, #first_step_range_end
8, #rest_step_range_start
10, #rest_step_range_end
20.0, #lambda_ac
0.055, #lambda_kl
False #noise_correction
],
[
"example_images/kitten.jpg", #input_image
"A kitten is sitting in a basket on a branch", #src_prompt
"a dog is sitting in a basket on a branch", #tgt_prompt
1.0, #edit_cfg
9, #number_of_renoising_iterations
1.0, #inersion_strength
True, #avg_gradients
0, #first_step_range_start
5, #first_step_range_end
8, #rest_step_range_start
10, #rest_step_range_end
20.0, #lambda_ac
0.055, #lambda_kl
False #noise_correction
],
[
"example_images/monkey.jpeg", #input_image
"a monkey sitting on a tree branch in the forest", #src_prompt
"a beaver sitting on a tree branch in the forest", #tgt_prompt
1.0, #edit_cfg
9, #number_of_renoising_iterations
1.0, #inersion_strength
True, #avg_gradients
0, #first_step_range_start
5, #first_step_range_end
8, #rest_step_range_start
10, #rest_step_range_end
20.0, #lambda_ac
0.055, #lambda_kl
True #noise_correction
],
[
"example_images/monkey.jpeg", #input_image
"a monkey sitting on a tree branch in the forest", #src_prompt
"a raccoon sitting on a tree branch in the forest", #tgt_prompt
1.0, #edit_cfg
9, #number_of_renoising_iterations
1.0, #inersion_strength
True, #avg_gradients
0, #first_step_range_start
5, #first_step_range_end
8, #rest_step_range_start
10, #rest_step_range_end
20.0, #lambda_ac
0.055, #lambda_kl
True #noise_correction
],
[
"example_images/lion.jpeg", #input_image
"a lion is sitting in the grass at sunset", #src_prompt
"a tiger is sitting in the grass at sunset", #tgt_prompt
1.0, #edit_cfg
9, #number_of_renoising_iterations
1.0, #inersion_strength
True, #avg_gradients
0, #first_step_range_start
5, #first_step_range_end
8, #rest_step_range_start
10, #rest_step_range_end
20.0, #lambda_ac
0.055, #lambda_kl
True #noise_correction
]
]
gr.Examples(examples=examples,
inputs=[
input_image,
src_prompt,
tgt_prompt,
edit_cfg,
number_of_renoising_iterations,
inersion_strength,
avg_gradients,
first_step_range_start,
first_step_range_end,
rest_step_range_start,
rest_step_range_end,
lambda_ac,
lambda_kl,
noise_correction
],
outputs=[
result
],
fn=main_pipeline,
cache_examples=True)
inputs = [
input_image,
src_prompt,
tgt_prompt,
edit_cfg,
number_of_renoising_iterations,
inersion_strength,
avg_gradients,
first_step_range_start,
first_step_range_end,
rest_step_range_start,
rest_step_range_end,
lambda_ac,
lambda_kl,
noise_correction
]
outputs = [
result
]
run_button.click(fn=main_pipeline, inputs=inputs, outputs=outputs)
demo.queue(max_size=50).launch(share=False) |