Spaces:
Running
on
Zero
Running
on
Zero
File size: 9,764 Bytes
aa8012e d1ecaf3 aa8012e d1ecaf3 aa8012e 3f80493 6643db0 3f80493 aa8012e 3f80493 aa8012e 3f80493 966cd3e 3f80493 d1ecaf3 3f80493 d559f02 3f80493 aa8012e 2f1a8cd aa8012e 2f1a8cd d1ecaf3 d06f403 aa8012e d1ecaf3 d06f403 aa8012e d1ecaf3 aa8012e e970ea2 d1ecaf3 e970ea2 d1ecaf3 aa8012e e970ea2 c7d22d1 d1ecaf3 e970ea2 c7d22d1 e970ea2 b8987a2 e970ea2 b8987a2 e970ea2 aa8012e 3f80493 b8987a2 d1ecaf3 b8987a2 aa8012e d1ecaf3 3f80493 d1ecaf3 aa8012e d06f403 |
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 |
import spaces
import time
import os
import gradio as gr
import torch
from einops import rearrange
from PIL import Image
from transformers import pipeline
from flux.cli import SamplingOptions
from flux.sampling import denoise, get_noise, get_schedule, prepare, unpack
from flux.util import load_ae, load_clip, load_flow_model, load_t5
from pulid.pipeline_flux import PuLIDPipeline
from pulid.utils import resize_numpy_image_long
# ํ์ ๋ฒ์ญ ๋ชจ๋ธ ๋ก๋
translator = pipeline("translation", model="Helsinki-NLP/opus-mt-ko-en")
def get_models(name: str, device: torch.device, offload: bool):
t5 = load_t5(device, max_length=128)
clip = load_clip(device)
model = load_flow_model(name, device="cpu" if offload else device)
model.eval()
ae = load_ae(name, device="cpu" if offload else device)
return model, ae, t5, clip
class FluxGenerator:
def __init__(self):
self.device = torch.device('cuda')
self.offload = False
self.model_name = 'flux-dev'
self.model, self.ae, self.t5, self.clip = get_models(
self.model_name,
device=self.device,
offload=self.offload,
)
self.pulid_model = PuLIDPipeline(self.model, 'cuda', weight_dtype=torch.bfloat16)
self.pulid_model.load_pretrain()
flux_generator = FluxGenerator()
@spaces.GPU
@torch.inference_mode()
def generate_image(
width,
height,
num_steps,
start_step,
guidance,
seed,
prompt,
id_image=None,
id_weight=1.0,
neg_prompt="",
true_cfg=1.0,
timestep_to_start_cfg=1,
max_sequence_length=128,
):
# ํ๊ธ ํ๋กฌํํธ๋ฅผ ์์ด๋ก ๋ฒ์ญ
if any('\u3131' <= c <= '\u318E' or '\uAC00' <= c <= '\uD7A3' for c in prompt):
translated = translator(prompt)[0]['translation_text']
prompt = translated
flux_generator.t5.max_length = max_sequence_length
seed = int(seed)
if seed == -1:
seed = None
opts = SamplingOptions(
prompt=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()
t0 = time.perf_counter()
use_true_cfg = abs(true_cfg - 1.0) > 1e-2
if id_image is not None:
id_image = resize_numpy_image_long(id_image, 1024)
id_embeddings, uncond_id_embeddings = flux_generator.pulid_model.get_id_embedding(id_image, cal_uncond=use_true_cfg)
else:
id_embeddings = None
uncond_id_embeddings = None
# prepare input
x = get_noise(
1,
opts.height,
opts.width,
device=flux_generator.device,
dtype=torch.bfloat16,
seed=opts.seed,
)
timesteps = get_schedule(
opts.num_steps,
x.shape[-1] * x.shape[-2] // 4,
shift=True,
)
if flux_generator.offload:
flux_generator.t5, flux_generator.clip = flux_generator.t5.to(flux_generator.device), flux_generator.clip.to(flux_generator.device)
inp = prepare(t5=flux_generator.t5, clip=flux_generator.clip, img=x, prompt=opts.prompt)
inp_neg = prepare(t5=flux_generator.t5, clip=flux_generator.clip, img=x, prompt=neg_prompt) if use_true_cfg else None
# offload TEs to CPU, load model to gpu
if flux_generator.offload:
flux_generator.t5, flux_generator.clip = flux_generator.t5.cpu(), flux_generator.clip.cpu()
torch.cuda.empty_cache()
flux_generator.model = flux_generator.model.to(flux_generator.device)
# denoise initial noise
x = denoise(
flux_generator.model, **inp, timesteps=timesteps, guidance=opts.guidance, id=id_embeddings, id_weight=id_weight,
start_step=start_step, uncond_id=uncond_id_embeddings, true_cfg=true_cfg,
timestep_to_start_cfg=timestep_to_start_cfg,
neg_txt=inp_neg["txt"] if use_true_cfg else None,
neg_txt_ids=inp_neg["txt_ids"] if use_true_cfg else None,
neg_vec=inp_neg["vec"] if use_true_cfg else None,
)
# offload model, load autoencoder to gpu
if flux_generator.offload:
flux_generator.model.cpu()
torch.cuda.empty_cache()
flux_generator.ae.decoder.to(x.device)
# decode latents to pixel space
x = unpack(x.float(), opts.height, opts.width)
with torch.autocast(device_type=flux_generator.device.type, dtype=torch.bfloat16):
x = flux_generator.ae.decode(x)
if flux_generator.offload:
flux_generator.ae.decoder.cpu()
torch.cuda.empty_cache()
t1 = time.perf_counter()
# bring into PIL format
x = x.clamp(-1, 1)
x = rearrange(x[0], "c h w -> h w c")
img = Image.fromarray((127.5 * (x + 1.0)).cpu().byte().numpy())
return img, str(opts.seed), flux_generator.pulid_model.debug_img_list
css = """
footer {
visibility: hidden;
}
"""
def create_demo(args, model_name: str, device: str = "cuda" if torch.cuda.is_available() else "cpu",
offload: bool = False):
with gr.Blocks(theme="Nymbo/Nymbo_Theme", css=css) as demo:
gr.Markdown("# ์ ๋ชฉ")
gr.Markdown("## ์ค๋ช
")
with gr.Row():
with gr.Column():
prompt = gr.Textbox(label="ํ๋กฌํํธ", value="์ด์ํ, ์๊ฐ, ์ํ์ ")
id_image = gr.Image(label="ID ์ด๋ฏธ์ง", source="webcam", type="numpy")
generate_btn = gr.Button("์์ฑ")
with gr.Column():
output_image = gr.Image(label="์์ฑ๋ ์ด๋ฏธ์ง")
with gr.Row():
with gr.Column():
gr.Markdown("## ์์")
all_examples = [
['์ฌ์๊ฐ \"PuLID for FLUX\"๋ผ๊ณ ์ฐ์ธ ๋น๋๋ ๋
น์ ํ์งํ์ ๋ค๊ณ ์๋ค', 'example_inputs/liuyifei.png'],
['์๋ชจ์ต ์ด์ํ', 'example_inputs/liuyifei.png'],
['VR ๊ธฐ์ ๋ถ์๊ธฐ์ ํฐ ๋จธ๋ฆฌ ์ฌ์ฑ', 'example_inputs/liuyifei.png'],
['์ด๋ฆฐ ์์ด๊ฐ ์์ด์คํฌ๋ฆผ์ ๋จน๊ณ ์๋ค', 'example_inputs/liuyifei.png'],
['๋จ์๊ฐ \"PuLID for FLUX\"๋ผ๊ณ ์ฐ์ธ ํ์งํ์ ๋ค๊ณ ์๋ค, ๊ฒจ์ธ, ๋ ๋ด๋ฆผ, ์ฐ ์ ์', 'example_inputs/pengwei.jpg'],
['์ด์ํ, ์ด๋ถ ์กฐ๋ช
', 'example_inputs/pengwei.jpg'],
['25์ธ ๋จ์ฑ์ ์ด๋์ด ํ๋กํ ์ฌ์ง, ์
์์ ์ฐ๊ธฐ๊ฐ ๋์ค๊ณ ์์', 'example_inputs/pengwei.jpg'],
['๋ฏธ๊ตญ ๋งํ ์คํ์ผ, ์๋
1๋ช
', 'example_inputs/pengwei.jpg'],
['์ด์ํ, ํฝ์ฌ ์คํ์ผ', 'example_inputs/pengwei.jpg'],
['์ด์ํ, ์ผ์ ์กฐ๊ฐ์', 'example_inputs/lecun.jpg'],
]
example_images = [example[1] for example in all_examples]
example_captions = [example[0] for example in all_examples]
gallery = gr.Gallery(
value=list(zip(example_images, example_captions)),
label="์์ ๊ฐค๋ฌ๋ฆฌ",
show_label=False,
elem_id="gallery",
columns=5,
rows=2,
object_fit="contain",
height="auto"
)
def fill_example(evt: gr.SelectData):
return [all_examples[evt.index][i] for i in [0, 1]]
gallery.select(
fill_example,
None,
[prompt, id_image],
)
generate_btn.click(
fn=generate_image,
inputs=[
gr.Slider(256, 1536, 896, step=16, visible=False), # width
gr.Slider(256, 1536, 1152, step=16, visible=False), # height
gr.Slider(1, 20, 20, step=1, visible=False), # num_steps
gr.Slider(0, 10, 0, step=1, visible=False), # start_step
gr.Slider(1.0, 10.0, 4, step=0.1, visible=False), # guidance
gr.Textbox(-1, visible=False), # seed
prompt,
id_image,
gr.Slider(0.0, 3.0, 1, step=0.05, visible=False), # id_weight
gr.Textbox("์ ํ์ง, ์ต์
์ ํ์ง, ํ
์คํธ, ์๋ช
, ์ํฐ๋งํฌ, ์ฌ๋ถ์ ํ๋ค๋ฆฌ", visible=False), # neg_prompt
gr.Slider(1.0, 10.0, 1, step=0.1, visible=False), # true_cfg
gr.Slider(0, 20, 1, step=1, visible=False), # timestep_to_start_cfg
gr.Slider(128, 512, 128, step=128, visible=False), # max_sequence_length
],
outputs=[output_image],
)
return demo
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser(description="PuLID for FLUX.1-dev")
parser.add_argument("--name", type=str, default="flux-dev", choices=list('flux-dev'),
help="ํ์ฌ๋ flux-dev๋ง ์ง์ํฉ๋๋ค")
parser.add_argument("--device", type=str, default="cuda" if torch.cuda.is_available() else "cpu",
help="์ฌ์ฉํ ๋๋ฐ์ด์ค")
parser.add_argument("--offload", action="store_true", help="์ฌ์ฉํ์ง ์์ ๋ ๋ชจ๋ธ์ CPU๋ก ์ฎ๊น๋๋ค")
parser.add_argument("--port", type=int, default=8080, help="์ฌ์ฉํ ํฌํธ")
parser.add_argument("--dev", action='store_true', help="๊ฐ๋ฐ ๋ชจ๋")
parser.add_argument("--pretrained_model", type=str, help='๊ฐ๋ฐ์ฉ')
args = parser.parse_args()
import huggingface_hub
huggingface_hub.login(os.getenv('HF_TOKEN'))
demo = create_demo(args, args.name, args.device, args.offload)
demo.launch() |