Spaces:
Running
on
Zero
Running
on
Zero
File size: 9,516 Bytes
4cc901a aca4c0c 4cc901a 0c67c24 4cc901a 6612f88 4cc901a 6fb545f c633291 6fb545f 1b839c2 6fb545f 4cc901a 6fb545f 4cc901a 6fb545f 4cc901a 6fb545f 4cc901a 6fb545f 748300c 1b839c2 4cc901a 6fb545f 4cc901a 6fb545f 4cc901a 6fb545f 4cc901a 6fb545f 1e3cd91 6fb545f 4cc901a 6fb545f 748300c 6fb545f 1e3cd91 1b839c2 6fb545f 4cc901a 6fb545f 4cc901a 6fb545f 4cc901a 6fb545f 4cc901a 6fb545f 1e3cd91 6fb545f 4cc901a 6fb545f b18d10b 6fb545f d30934f 6fb545f 4cc901a a7b8414 4cc901a 062e9dc 6f80cb0 4cc901a 062e9dc 6f80cb0 31c7137 6f80cb0 062e9dc 08b018a 062e9dc 6f80cb0 f279f58 062e9dc 4cc901a 2e0ad47 4cc901a 6f80cb0 4cc901a 6fb545f 4cc901a 6fb545f 779331d 6fb545f 779331d 6fb545f |
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 |
import os
import re
import time
from io import BytesIO
import uuid
from dataclasses import dataclass
from glob import iglob
import argparse
from einops import rearrange
from fire import Fire
from PIL import ExifTags, Image
import spaces
import torch
import torch.nn.functional as F
import gradio as gr
import numpy as np
from transformers import pipeline
from flux.sampling import denoise, get_schedule, prepare, unpack
from flux.util import (configs, embed_watermark, load_ae, load_clip, load_flow_model, load_t5)
from huggingface_hub import login
login(token=os.getenv('Token'))
import torch
@dataclass
class SamplingOptions:
source_prompt: str
target_prompt: str
# prompt: str
width: int
height: int
num_steps: int
guidance: float
seed: int | None
@torch.inference_mode()
def encode(init_image, torch_device):
init_image = torch.from_numpy(init_image).permute(2, 0, 1).float() / 127.5 - 1
init_image = init_image.unsqueeze(0)
init_image = init_image.to(torch_device)
with torch.no_grad():
init_image = ae.encode(init_image.to()).to(torch.bfloat16)
return init_image
device = "cuda" if torch.cuda.is_available() else "cpu"
name = 'flux-dev'
ae = load_ae(name, device)
t5 = load_t5(device, max_length=256 if name == "flux-schnell" else 512)
clip = load_clip(device)
model = load_flow_model(name, device=device)
offload = False
name = "flux-dev"
is_schnell = False
feature_path = 'feature'
output_dir = 'result'
add_sampling_metadata = True
@spaces.GPU(duration=120)
@torch.inference_mode()
def edit(init_image, source_prompt, target_prompt, num_steps, inject_step, guidance, seed):
device = "cuda" if torch.cuda.is_available() else "cpu"
torch.cuda.empty_cache()
seed = None
shape = init_image.shape
new_h = shape[0] if shape[0] % 16 == 0 else shape[0] - shape[0] % 16
new_w = shape[1] if shape[1] % 16 == 0 else shape[1] - shape[1] % 16
init_image = init_image[:new_h, :new_w, :]
width, height = init_image.shape[0], init_image.shape[1]
init_image = torch.from_numpy(init_image).permute(2, 0, 1).float() / 127.5 - 1
init_image = init_image.unsqueeze(0)
init_image = init_image.to(device)
with torch.no_grad():
init_image = ae.encode(init_image.to()).to(torch.bfloat16)
print(init_image.shape)
rng = torch.Generator(device="cpu")
opts = SamplingOptions(
source_prompt=source_prompt,
target_prompt=target_prompt,
width=width,
height=height,
num_steps=num_steps,
guidance=guidance,
seed=seed,
)
if opts.seed is None:
opts.seed = torch.Generator(device="cpu").seed()
print(f"Generating with seed {opts.seed}:\n{opts.source_prompt}")
t0 = time.perf_counter()
opts.seed = None
#############inverse#######################
info = {}
info['feature'] = {}
info['inject_step'] = inject_step
with torch.no_grad():
inp = prepare(t5, clip, init_image, prompt=opts.source_prompt)
inp_target = prepare(t5, clip, init_image, prompt=opts.target_prompt)
timesteps = get_schedule(opts.num_steps, inp["img"].shape[1], shift=(name != "flux-schnell"))
# inversion initial noise
with torch.no_grad():
z, info = denoise(model, **inp, timesteps=timesteps, guidance=1, inverse=True, info=info)
inp_target["img"] = z
timesteps = get_schedule(opts.num_steps, inp_target["img"].shape[1], shift=(name != "flux-schnell"))
# denoise initial noise
x, _ = denoise(model, **inp_target, timesteps=timesteps, guidance=guidance, inverse=False, info=info)
# decode latents to pixel space
x = unpack(x.float(), opts.width, opts.height)
output_name = os.path.join(output_dir, "img_{idx}.jpg")
if not os.path.exists(output_dir):
os.makedirs(output_dir)
idx = 0
else:
fns = [fn for fn in iglob(output_name.format(idx="*")) if re.search(r"img_[0-9]+\.jpg$", fn)]
if len(fns) > 0:
idx = max(int(fn.split("_")[-1].split(".")[0]) for fn in fns) + 1
else:
idx = 0
device = torch.device("cuda")
with torch.autocast(device_type=device.type, dtype=torch.bfloat16):
x = ae.decode(x)
if torch.cuda.is_available():
torch.cuda.synchronize()
t1 = time.perf_counter()
fn = output_name.format(idx=idx)
print(f"Done in {t1 - t0:.1f}s. Saving {fn}")
# bring into PIL format and save
x = x.clamp(-1, 1)
x = embed_watermark(x.float())
x = rearrange(x[0], "c h w -> h w c")
img = Image.fromarray((127.5 * (x + 1.0)).cpu().byte().numpy())
exif_data = Image.Exif()
exif_data[ExifTags.Base.Software] = "AI generated;txt2img;flux"
exif_data[ExifTags.Base.Make] = "Black Forest Labs"
exif_data[ExifTags.Base.Model] = name
if add_sampling_metadata:
exif_data[ExifTags.Base.ImageDescription] = source_prompt
# img.save(fn, exif=exif_data, quality=95, subsampling=0)
print("End Edit")
return img
def create_demo(model_name: str, device: str = "cuda:0" if torch.cuda.is_available() else "cpu", offload: bool = False):
is_schnell = model_name == "flux-schnell"
title = r"""
<h1 align="center">Taming Rectified Flow for Inversion and Editing</h1>
"""
description = r"""
<b>Official 🤗 Gradio demo</b> for <a href='https://github.com/wangjiangshan0725/RF-Solver-Edit' target='_blank'><b>Taming Rectified Flow for Inversion and Editing</b></a>.<br>
❗️❗️❗️[<b>Important</b>] Editing steps:<br>
1️⃣ Upload images you want to edit (The resolution is expected be less than 1360*768, or the memory of GPU may be not enough.) <br>
2️⃣ Enter the source prompt, which describes the content of the image you unload. The source prompt is not mandatory; you can also leave it to null. <br>
3️⃣ Enter the target prompt which describes the content of the expected images after editing. <br>
4️⃣ Click the <b>Generate</b> button to start editing. <br>
5️⃣ We suggest to adjust the value of **feature sharing steps** for better results.<br>
"""
article = r"""
If our work is helpful, please help to ⭐ the <a href='https://github.com/wangjiangshan0725/RF-Solver-Edit' target='_blank'>Github Repo</a>. Thanks!
[![GitHub Stars](https://img.shields.io/github/stars/wangjiangshan0725/RF-Solver-Edit?style=social)](https://github.com/wangjiangshan0725/RF-Solver-Edit)
"""
css = '''
.gradio-container {width: 85% !important}
'''
with gr.Blocks(css=css) as demo:
# gr.Markdown(f"# Official Demo for Taming Rectified Flow for Inversion and Editing")
gr.Markdown(title)
gr.Markdown(description)
gr.Markdown(article)
with gr.Row():
with gr.Column():
source_prompt = gr.Textbox(label="Source Prompt", value="")
target_prompt = gr.Textbox(label="Target Prompt", value="")
# source_prompt = gr.Text(
# label="Source Prompt",
# show_label=False,
# max_lines=1,
# placeholder="Enter your source prompt",
# container=False,
# value=""
# )
# target_prompt = gr.Text(
# label="Target Prompt",
# show_label=False,
# max_lines=1,
# placeholder="Enter your target prompt",
# container=False,
# value=""
# )
init_image = gr.Image(label="Input Image", visible=True)
generate_btn = gr.Button("Generate")
with gr.Column():
with gr.Accordion("Advanced Options", open=True):
num_steps = gr.Slider(1, 30, 25, step=1, label="Total timesteps")
inject_step = gr.Slider(1, 15, 3, step=1, label="Feature sharing steps")
guidance = gr.Slider(1.0, 10.0, 2, step=0.1, label="Guidance", interactive=not is_schnell)
# seed = gr.Textbox(0, label="Seed (-1 for random)", visible=False)
# add_sampling_metadata = gr.Checkbox(label="Add sampling parameters to metadata?", value=False)
output_image = gr.Image(label="Generated Image")
generate_btn.click(
fn=edit,
inputs=[init_image, source_prompt, target_prompt, num_steps, inject_step, guidance],
outputs=[output_image]
)
return demo
# if __name__ == "__main__":
# import argparse
# parser = argparse.ArgumentParser(description="Flux")
# parser.add_argument("--name", type=str, default="flux-dev", choices=list(configs.keys()), help="Model name")
# parser.add_argument("--device", type=str, default="cuda:0" if torch.cuda.is_available() else "cpu", help="Device to use")
# parser.add_argument("--offload", action="store_true", help="Offload model to CPU when not in use")
# parser.add_argument("--share", action="store_true", help="Create a public link to your demo")
# parser.add_argument("--port", type=int, default=41035)
# args = parser.parse_args()
demo = create_demo("flux-dev", "cuda")
demo.launch() |