Spaces:
Running
on
Zero
Running
on
Zero
File size: 11,723 Bytes
336dbcf d7a1a69 d4ee1f3 336dbcf d7a1a69 c46f4d1 1f31600 336dbcf d00ce10 52b4b13 336dbcf 02b6647 801b180 02b6647 d7a1a69 336dbcf 22406d6 f0190bb 22406d6 4cc8081 336dbcf 1355f82 336dbcf 02b6647 336dbcf 02b6647 801b180 02b6647 801b180 02b6647 336dbcf 3de5f1f 801b180 336dbcf 3446475 336dbcf b8381d1 336dbcf b8381d1 336dbcf b8381d1 336dbcf b8381d1 1f31600 b8381d1 336dbcf b8381d1 336dbcf f2d2079 b8381d1 f2d2079 336dbcf b8381d1 336dbcf b8381d1 |
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 |
import gradio as gr
import torch
from diffusers import FluxPipeline, StableDiffusion3Pipeline
from PIL import Image
from typing import Optional
import os
import random
import numpy as np
import spaces
import huggingface_hub
import copy
from FlowEdit_utils import FlowEditSD3, FlowEditFLUX
SD3STRING = 'stabilityai/stable-diffusion-3-medium-diffusers'
FLUXSTRING = 'black-forest-labs/FLUX.1-dev'
device = "cuda" if torch.cuda.is_available() else "cpu"
# device = "cpu"
# model_type = 'SD3'
pipe_sd3 = StableDiffusion3Pipeline.from_pretrained("stabilityai/stable-diffusion-3-medium-diffusers", torch_dtype=torch.float16, token=os.getenv('HF_ACCESS_TOK'))
pipe_flux = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.float16, token=os.getenv('HF_ACCESS_TOK'))
# pipe_sd3.to(device)
# pipe_flux.to(device)
# scheduler = pipe.scheduler
# pipe = pipe.to(device)
loaded_model = 'None'
def on_model_change(model_type):
if model_type == 'SD3':
T_steps_value = 50
src_guidance_scale_value = 3.5
tar_guidance_scale_value = 13.5
n_max_value = 33
elif model_type == 'FLUX':
T_steps_value = 28
src_guidance_scale_value = 1.5
tar_guidance_scale_value = 5.5
n_max_value = 24
else:
raise NotImplementedError(f"Model type {model_type} not implemented")
return T_steps_value, src_guidance_scale_value, tar_guidance_scale_value, n_max_value
def get_examples():
case = [
["inputs/cat.png", "SD3", 50, 3.5, 13.5, 33, "A small, fluffy kitten sitting in a grassy field. The kitten is positioned in the center of the scene, surrounded by a field. The kitten appears to be looking at something in the field.", "A small puppy sitting in a grassy field. The puppy is positioned in the center of the scene, surrounded by a field. The puppy appears to be looking at something in the field.", 0, 1, 42, "example_outs/cat_puppy_sd3.png"],
["inputs/iguana.png", "SD3", 50, 3.5, 13.5, 31, "A large orange lizard sitting on a rock near the ocean. The lizard is positioned in the center of the scene, with the ocean waves visible in the background. The rock is located close to the water, providing a picturesque setting for the lizard''s resting spot.", "A large dragon sitting on a rock near the ocean. The dragon is positioned in the center of the scene, with the ocean waves visible in the background. The rock is located close to the water, providing a picturesque setting for the dragon''s resting spot.", 0, 1, 42, "example_outs/iguana_dragon.png"],
["inputs/cat.png", "FLUX", 28, 1.5, 5.5, 23, "A small, fluffy kitten sitting in a grassy field. The kitten is positioned in the center of the scene, surrounded by a field. The kitten appears to be looking at something in the field.", "A small puppy sitting in a grassy field. The puppy is positioned in the center of the scene, surrounded by a field. The puppy appears to be looking at something in the field.", 0, 1, 42, "example_outs/cat_puppy_flux.png"],
["inputs/gas_station.png", "FLUX", 28, 1.5, 5.5, 23, "A gas station with a white and red sign that reads \"CAFE\" There are several cars parked in front of the gas station, including a white car and a van.", "A gas station with a white and red sign that reads \"LOVE\" There are several cars parked in front of the gas station, including a white car and a van.", 0, 1, 42, "example_outs/gas_cafe_love.png"],
["inputs/steak.png", "FLUX", 28, 1.5, 5.5, 23, "A steak accompanied by a side of leaf salad.", "A bread roll accompanied by a side of leaf salad.", 0, 1, 42, "example_outs/steak_bread.png"],
["inputs/kill_bill.png", "FLUX", 28, 2.5, 6.5, 22, "a blonde woman in a yellow jumpsuit holding a sword in front of her face", "a blonde woman in a yellow jumpsuit holding a sword in front of her face, anime style drawing", 14, 1, 42, "example_outs/kill_bill_anime.png"],
]
return case
@spaces.GPU(duration=75)
def FlowEditRun(
image_src: str,
model_type: str,
T_steps: int,
src_guidance_scale: float,
tar_guidance_scale: float,
n_max: int,
src_prompt: str,
tar_prompt: str,
n_min: int,
n_avg: int,
seed: int,
):
if not len(src_prompt):
raise gr.Error("source prompt cannot be empty")
if not len(tar_prompt):
raise gr.Error("target prompt cannot be empty")
# global pipe_sd3
# global scheduler
# global loaded_model
# reload model only if different from the loaded model
# if loaded_model != model_type:
if model_type == 'FLUX':
# pipe = FluxPipeline.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.float16, token=os.getenv('HF_ACCESS_TOK'))
pipe = pipe_flux.to(device)
elif model_type == 'SD3':
# pipe = StableDiffusion3Pipeline.from_pretrained("stabilityai/stable-diffusion-3-medium-diffusers", torch_dtype=torch.float16, token=os.getenv('HF_ACCESS_TOK'))
pipe = pipe_sd3.to(device)
else:
raise NotImplementedError(f"Model type {model_type} not implemented")
scheduler = pipe.scheduler
# pipe = pipe.to(device)
# set seed
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
# load image
image = Image.open(image_src)
# crop image to have both dimensions divisibe by 16 - avoids issues with resizing
image = image.crop((0, 0, image.width - image.width % 16, image.height - image.height % 16))
image_src = pipe.image_processor.preprocess(image)
# image_tar = pipe.image_processor.postprocess(image_src)
# return image_tar[0]
# cast image to half precision
image_src = image_src.to(device).half()
with torch.autocast("cuda"), torch.inference_mode():
x0_src_denorm = pipe.vae.encode(image_src).latent_dist.mode()
x0_src = (x0_src_denorm - pipe.vae.config.shift_factor) * pipe.vae.config.scaling_factor
# send to cuda
x0_src = x0_src.to(device)
negative_prompt = "" # optionally add support for negative prompts (SD3)
if model_type == 'SD3':
x0_tar = FlowEditSD3(pipe,
scheduler,
x0_src,
src_prompt,
tar_prompt,
negative_prompt,
T_steps,
n_avg,
src_guidance_scale,
tar_guidance_scale,
n_min,
n_max,)
elif model_type == 'FLUX':
x0_tar = FlowEditFLUX(pipe,
scheduler,
x0_src,
src_prompt,
tar_prompt,
negative_prompt,
T_steps,
n_avg,
src_guidance_scale,
tar_guidance_scale,
n_min,
n_max,)
else:
raise NotImplementedError(f"Sampler type {model_type} not implemented")
x0_tar_denorm = (x0_tar / pipe.vae.config.scaling_factor) + pipe.vae.config.shift_factor
with torch.autocast("cuda"), torch.inference_mode():
image_tar = pipe.vae.decode(x0_tar_denorm, return_dict=False)[0]
image_tar = pipe.image_processor.postprocess(image_tar)
return image_tar[0]
# title = "FlowEdit: Inversion-Free Text-Based Editing Using Pre-Trained Flow Models"
intro = """
<h1 style="font-weight: 1000; text-align: center; margin: 0px;">FlowEdit: Inversion-Free Text-Based Editing Using Pre-Trained Flow Models</h1>
<h3 style="margin-bottom: 10px; text-align: center;">
<a href="https://arxiv.org/abs/2412.08629">[Paper]</a> |
<a href="https://matankleiner.github.io/flowedit/">[Project Page]</a> |
<a href="https://github.com/fallenshock/FlowEdit">[Code]</a>
</h3>
<br> π¨ Edit your image using Flow models! upload an image, add a description of it, and specify the edits you want to make
<h3>Notes:</h3>
<ol>
<li>We use FLUX.1 dev and SD3 for the demo. The models are large and may take a while to load.</li>
<li>We recommend 1024x1024 images for the best results. If the input images are too large, there may be out-of-memory errors.</li>
<li>Default hyperparameters for each model used in the paper are provided as examples. Feel free to experiment with them as well.</li>
</ol>
"""
# article = """
# π **Citation**
# ```bibtex
# @article{aaa,
# author = {},
# title = {},
# journal = {},
# year = {2024},
# url = {}
# }
# ```
# """
css="""
#col-container {
margin: 0 auto;
max-width: 960px;
}
"""
with gr.Blocks(css=css) as demo:
with gr.Column(elem_id="col-container"):
gr.HTML(intro)
# with gr.Row():
# gr.LoginButton(value="Login to HF (For SD3 and FLUX access)", variant="primary")
with gr.Row():
with gr.Column():
image_src = gr.Image(type="filepath", label="Source Image", value="inputs/cat.png",)
src_prompt = gr.Textbox(lines=2, label="Source Prompt", value="a cat sitting in the grass")
tar_prompt = gr.Textbox(lines=2, label="Target Prompt", value="a puppy sitting in the grass")
submit_button = gr.Button("Run FlowEdit", variant="primary")
with gr.Row():
model_type = gr.Dropdown(["SD3", "FLUX"], label="Model Type", value="SD3")
T_steps = gr.Slider(value=50, label="Total Steps", minimum=1, maximum=50)
n_max = gr.Slider(minimum=1, maximum=50, value=33, label="n_max", info="control the strength of the edit")
src_guidance_scale = gr.Slider(minimum=1.0, maximum=30.0, value=3.5, label="src_guidance_scale")
tar_guidance_scale = gr.Slider(minimum=1.0, maximum=30.0, value=13.5, label="tar_guidance_scale")
with gr.Column():
image_tar = gr.Image(label="Output", type="pil", show_label=True, format="png",)
with gr.Accordion(label="Advanced Settings", open=False):
# additional inputs
n_min = gr.Number(value=0, label="n_min (for improved style edits)")
n_avg = gr.Number(value=1, label="n_avg (improve structure at the cost of runtime)", minimum=1)
seed = gr.Number(value=42, label="seed")
submit_button.click(
fn=FlowEditRun,
inputs=[
image_src,
model_type,
T_steps,
src_guidance_scale,
tar_guidance_scale,
n_max,
src_prompt,
tar_prompt,
n_min,
n_avg,
seed,
],
outputs=[
image_tar,
],
)
gr.Examples(
label="Examples",
examples=get_examples(),
inputs=[image_src, model_type, T_steps, src_guidance_scale, tar_guidance_scale, n_max, src_prompt, tar_prompt, n_min, n_avg, seed, image_tar],
outputs=[image_tar],
)
model_type.input(fn=on_model_change, inputs=[model_type], outputs=[T_steps, src_guidance_scale, tar_guidance_scale, n_max])
# gr.HTML(article)
demo.queue()
demo.launch( ) |