File size: 14,106 Bytes
aa8012e
 
 
 
 
 
 
 
9084b6a
aa8012e
 
 
 
 
 
 
9084b6a
aa8012e
 
 
 
 
 
 
9084b6a
 
aa8012e
 
 
3f80493
 
6643db0
3f80493
9084b6a
3f80493
aa8012e
 
 
3f80493
 
 
 
 
 
 
 
966cd3e
3f80493
80465fd
 
3f80493
 
 
80465fd
 
 
 
3f80493
80465fd
3f80493
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
dd6fe43
3f80493
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9084b6a
 
 
 
 
 
aa8012e
 
 
 
 
 
e6c5ff5
aa8012e
e6c5ff5
aa8012e
3efa982
80465fd
e6c5ff5
aa8012e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1a4216e
aa8012e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
80465fd
aa8012e
 
 
 
80465fd
aa8012e
 
 
 
80465fd
aa8012e
 
 
 
80465fd
aa8012e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
80465fd
 
aa8012e
 
 
 
 
80465fd
aa8012e
 
 
80465fd
 
aa8012e
 
3f80493
80465fd
 
aa8012e
 
 
 
 
 
 
 
 
 
 
 
3f80493
 
aa8012e
 
 
 
 
 
 
 
 
 
80465fd
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
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

NSFW_THRESHOLD = 0.85

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)
    nsfw_classifier = pipeline("image-classification", model="Falconsai/nsfw_image_detection", device=device)
    return model, ae, t5, clip, nsfw_classifier


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, self.nsfw_classifier = 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(
        prompt,
        id_image,
        start_step,
        guidance,
        seed,
        true_cfg,
        width=896,
        height=1152,
        num_steps=20,
        id_weight=1.0,
        neg_prompt="bad quality, worst quality, text, signature, watermark, extra limbs",
        timestep_to_start_cfg=1,
        max_sequence_length=128,
):
    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()
    print(f"Generating '{opts.prompt}' with seed {opts.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()

    print(f"Done in {t1 - t0:.1f}s.")
    # bring into PIL format
    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())
    nsfw_score = [x["score"] for x in flux_generator.nsfw_classifier(img) if x["label"] == "nsfw"][0]
    if nsfw_score < NSFW_THRESHOLD:
        return img, str(opts.seed), flux_generator.pulid_model.debug_img_list
    else:
        return (None, f"Your generated image may contain NSFW (with nsfw_score: {nsfw_score}) content",
                flux_generator.pulid_model.debug_img_list)

_HEADER_ = '''
<div style="text-align: center; max-width: 650px; margin: 0 auto;">
    <h1 style="font-size: 2.5rem; font-weight: 700; margin-bottom: 1rem; display: contents;">PuLID for FLUX</h1>
    <p style="font-size: 1rem; margin-bottom: 1.5rem;">Paper: <a href='https://arxiv.org/abs/2404.16022' target='_blank'>PuLID: Pure and Lightning ID Customization via Contrastive Alignment</a> | Codes: <a href='https://github.com/ToTheBeginning/PuLID' target='_blank'>GitHub</a></p>
</div>

❗️❗️❗️**Tips:**

- `timestep to start inserting ID:` The smaller the value, the higher the fidelity, but the lower the editability; the higher the value, the lower the fidelity, but the higher the editability. **The recommended range for this value is between 0 and 4**. For photorealistic scenes, we recommend using 4; for stylized scenes, we recommend using 0-1. If you are not satisfied with the similarity, you can lower this value; conversely, if you are not satisfied with the editability, you can increase this value.
- `true CFG scale:` In most scenarios, it is recommended to use a fake CFG, i.e., setting the true CFG scale to 1, and just adjusting the guidance scale. This is also more efficiency. However, in a few cases, utilizing a true CFG can yield better results. For more detaileds, please refer to the [doc](https://github.com/ToTheBeginning/PuLID/blob/main/docs/pulid_for_flux.md#useful-tips).
- `Learn more about the model:` please refer to the <a href='https://github.com/ToTheBeginning/PuLID/blob/main/docs/pulid_for_flux.md' target='_blank'>github doc</a> for more details and info about the model, we provide the detail explanation about the above two parameters in the doc.
- `Examples:` we provide some examples (we have cached them, so just click them to see what the model can do) in the bottom, you can try these example prompts first
'''  # noqa E501

_CITE_ = r"""
If PuLID is helpful, please help to ⭐ the <a href='https://github.com/ToTheBeginning/PuLID' target='_blank'> Github Repo</a>. Thanks!
---
📧 **Contact**
If you have any questions or feedbacks, feel free to open a discussion or contact <b>[email protected]</b>.
"""  # noqa E501


def create_demo(args, model_name: str, device: str = "cuda" if torch.cuda.is_available() else "cpu",
                offload: bool = False):
    with gr.Blocks() as demo:
        gr.Markdown(_HEADER_)

        with gr.Row():
            with gr.Column():
                prompt = gr.Textbox(label="Prompt", value="portrait, color, cinematic")
                id_image = gr.Image(label="ID Image")
                id_weight = gr.Slider(0.0, 3.0, 1, step=0.05, label="id weight")

                width = gr.Slider(256, 1536, 896, step=16, label="Width")
                height = gr.Slider(256, 1536, 1152, step=16, label="Height")
                num_steps = gr.Slider(1, 20, 20, step=1, label="Number of steps")
                start_step = gr.Slider(0, 10, 0, step=1, label="timestep to start inserting ID")
                guidance = gr.Slider(1.0, 10.0, 4, step=0.1, label="Guidance")
                seed = gr.Textbox(-1, label="Seed (-1 for random)")
                max_sequence_length = gr.Slider(128, 512, 128, step=128,
                                                label="max_sequence_length for prompt (T5), small will be faster")

                with gr.Accordion("Advanced Options (True CFG, true_cfg_scale=1 means use fake CFG, >1 means use true CFG, if using true CFG, we recommend set the guidance scale to 1)", open=False):    # noqa E501
                    neg_prompt = gr.Textbox(
                        label="Negative Prompt",
                        value="bad quality, worst quality, text, signature, watermark, extra limbs")
                    true_cfg = gr.Slider(1.0, 10.0, 1, step=0.1, label="true CFG scale")
                    timestep_to_start_cfg = gr.Slider(0, 20, 1, step=1, label="timestep to start cfg", visible=args.dev)

                generate_btn = gr.Button("Generate")

            with gr.Column():
                output_image = gr.Image(label="Generated Image", format='png')
                seed_output = gr.Textbox(label="Used Seed")
                intermediate_output = gr.Gallery(label='Output', elem_id="gallery", visible=args.dev)
                gr.Markdown(_CITE_)

        with gr.Row(), gr.Column():
                gr.Markdown("## Examples")
                example_inps = [
                    [
                        'a woman holding sign with glowing green text \"PuLID for FLUX\"',
                        'example_inputs/liuyifei.png',
                        4, 4, 2680261499100305976, 1
                    ],
                    [
                        'portrait, side view',
                        'example_inputs/liuyifei.png',
                        4, 4, 180825677246321775, 1
                    ],
                    [
                        'white-haired woman with vr technology atmosphere, revolutionary exceptional magnum with remarkable details',  # noqa E501
                        'example_inputs/liuyifei.png',
                        4, 4, 16942328329935464989, 1
                    ],
                    [
                        'a young child is eating Icecream',
                        'example_inputs/liuyifei.png',
                        4, 4, 4527590969012358757, 1
                    ],
                    [
                        'a man is holding a sign with text \"PuLID for FLUX\", winter, snowing, top of the mountain',
                        'example_inputs/pengwei.jpg',
                        4, 4, 6273700647573240909, 1
                    ],
                    [
                        'portrait, candle light',
                        'example_inputs/pengwei.jpg',
                        4, 4, 17522759474323955700, 1
                    ],
                    [
                        'profile shot dark photo of a 25-year-old male with smoke escaping from his mouth, the backlit smoke gives the image an ephemeral quality, natural face, natural eyebrows, natural skin texture, award winning photo, highly detailed face, atmospheric lighting, film grain, monochrome',  # noqa E501
                        'example_inputs/pengwei.jpg',
                        4, 4, 17733156847328193625, 1
                    ],
                    [
                        'American Comics, 1boy',
                        'example_inputs/pengwei.jpg',
                        1, 4, 13223174453874179686, 1
                    ],
                    [
                        'portrait, pixar',
                        'example_inputs/pengwei.jpg',
                        1, 4, 9445036702517583939, 1
                    ],
                ]
                gr.Examples(examples=example_inps, inputs=[prompt, id_image, start_step, guidance, seed, true_cfg],
                            label='fake CFG', cache_examples='lazy', outputs=[output_image, seed_output],
                            fn=generate_image)

                example_inps = [
                    [
                        'portrait, made of ice sculpture',
                        'example_inputs/lecun.jpg',
                        1, 1, 7717391560531186077, 5
                    ],
                ]
                gr.Examples(examples=example_inps, inputs=[prompt, id_image, start_step, guidance, seed, true_cfg],
                            label='true CFG', cache_examples='lazy', outputs=[output_image, seed_output],
                            fn=generate_image)

        generate_btn.click(
            fn=generate_image,
            inputs=[prompt, id_image, start_step, guidance, seed, true_cfg, width, height, num_steps, id_weight,
                    neg_prompt, timestep_to_start_cfg, max_sequence_length],
            outputs=[output_image, seed_output, intermediate_output],
        )

    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="currently only support flux-dev")
    parser.add_argument("--device", type=str, default="cuda" 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("--port", type=int, default=8080, help="Port to use")
    parser.add_argument("--dev", action='store_true', help="Development mode")
    parser.add_argument("--pretrained_model", type=str, help='for development')
    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()