File size: 12,276 Bytes
19ed964
 
 
 
 
 
 
 
 
 
 
 
c244c09
19ed964
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c937232
19ed964
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6f1c8f2
19ed964
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
6f1c8f2
19ed964
1fbb135
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
19ed964
 
 
 
bb56969
1fbb135
19ed964
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1fbb135
 
66b3640
 
 
 
bb56969
66b3640
 
 
 
 
 
 
 
 
bb56969
66b3640
 
 
 
 
 
 
 
 
bb56969
66b3640
 
 
 
 
 
 
 
 
 
 
 
 
 
bb56969
66b3640
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
19ed964
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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 flux.details 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 eva_clip.model_configs.fluxpipeline import ToonMagePipeline
from toonmage.utils import resize_numpy_image_long


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.toonmage_model = ToonMagePipeline(self.model, 'cuda', weight_dtype=torch.bfloat16)
        self.toonmage_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,
):
    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.toonmage_model.get_id_embedding(id_image, cal_uncond=use_true_cfg)
    else:
        id_embeddings = None
        uncond_id_embeddings = None

    print(id_embeddings)

    # prepare input
    x = get_noise(
        1,
        opts.height,
        opts.width,
        device=flux_generator.device,
        dtype=torch.bfloat16,
        seed=opts.seed,
    )
    print(x)
    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())
    return img, str(opts.seed), flux_generator.toonmage_model.debug_img_list

MARKDOWN = """
This demo utilizes <a href="https://huggingface.co/black-forest-labs/FLUX.1-dev">FLUX Pipeline</a> for Image to Image Translation

**Tips**

- Smaller value of timestep to start inserting ID would lead to higher fidelity, however, it will reduce the editability; and vice versa.
Its value range is from 0 - 4. If you want to generate a stylized scene; use the value of 0 - 1. If you want to generate a photorealistic image; use the value of 4. 

-It is recommended to use fake CFG by setting the true CFG scale value to 1 while you can vary the guidance scale. However, in a few cases, utilizing a true CFG can yield better results.

Try out with different prompts using your image and do provide your feedback.

**Demo by [Sunder Ali Khowaja](https://sander-ali.github.io) - [X](https://x.com/SunderAKhowaja) -[Github](https://github.com/sander-ali) -[Hugging Face](https://huggingface.co/SunderAli17)**
"""

theme = gr.themes.Soft(
    font=[gr.themes.GoogleFont('Source Code Pro'), gr.themes.GoogleFont('Public Sans'), 'system-ui', 'sans-serif'],
)
js_func = """
function refresh() {
    const url = new URL(window.location);
    if (url.searchParams.get('__theme') !== 'dark') {
        url.searchParams.set('__theme', 'dark');
        window.location.href = url.href;
    }
}
"""


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

        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")
                seed_output = gr.Textbox(label="Used Seed")
                intermediate_output = gr.Gallery(label='Output', elem_id="gallery", visible=args.dev)


        with gr.Row(), gr.Column():
                gr.Markdown("## Examples")
                example_inps = [
                    [
                        'a high quality digital avatar, eating icecream',
                        'sample_img/image1.png',
                        4, 4, 2680261499100305976, 1
                    ],
                    [
                        'portrait, side view',
                        'sample_img/image1.png',
                        4, 4, 1205240166692517553, 1
                    ],
                    [
                        'white-haired man with vr technology', 
                        'sample_img/image1.png',
                        4, 4, 6349424134217931066, 1
                    ],
                    [
                        'a young child is eating Icecream',
                        'sample_img/image1.png',
                        4, 4, 10606046113565776207, 1
                    ],
                    [
                        'a man is holding Pakistani flag on top of a mountain',
                        'sample_img/test1.jpg',
                        4, 4, 2410129802683836089, 1
                    ],
                    [
                        'portrait, candle light',
                        'sample_img/test1.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
                        'sample_img/test1.jpg',
                        4, 4, 17733156847328193625, 1
                    ],
                    [
                        'DC comics, Flash',
                        'sample_img/test1.jpg',
                        1, 4, 13223174453874179686, 1
                    ],
                    [
                        'portrait, pixar',
                        'sample_img/test1.jpg',
                        1, 4, 9445036702517583939, 1
                    ],
                ]
                gr.Examples(examples=example_inps, inputs=[prompt, id_image, start_step, guidance, seed, true_cfg],
                            label='fake CFG')

                example_inps = [
                    [
                        'portrait, made of ice sculpture',
                        'sample_img/test1.jpg',
                        1, 1, 3811899118709451814, 5
                    ],
                ]
                gr.Examples(examples=example_inps, inputs=[prompt, id_image, start_step, guidance, seed, true_cfg],
                            label='true CFG')


        generate_btn.click(
            fn=generate_image,
            inputs=[width, height, num_steps, start_step, guidance, seed, prompt, id_image, id_weight, neg_prompt,
                    true_cfg, 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="ToonMage with FLUX")
    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()