Update app.py

app.py

@@ -1,225 +1,320 @@
-# import spaces
+import spaces
+import time
+import os
+
 import gradio as gr
-import numpy as np
 import torch
-
-from pulid import attention_processor as attention
-from pulid.pipeline import PuLIDPipeline
-from pulid.utils import resize_numpy_image_long, seed_everything
-
-torch.set_grad_enabled(True)
-
-pipeline = PuLIDPipeline()
-
-# other params
-DEFAULT_NEGATIVE_PROMPT = (
-    'flaws in the eyes, flaws in the face, flaws, lowres, non-HDRi, low quality, worst quality,'
-    'artifacts noise, text, watermark, glitch, deformed, mutated, ugly, disfigured, hands, '
-    'low resolution, partially rendered objects,  deformed or partially rendered eyes, '
-    'deformed, deformed eyeballs, cross-eyed,blurry'
-)
-
-
-# @spaces.GPU
-def run(*args):
-    id_image = args[0]
-    supp_images = args[1:4]
-    prompt, neg_prompt, scale, n_samples, seed, steps, H, W, id_scale, mode, id_mix = args[4:]
-
-    pipeline.debug_img_list = []
-    if mode == 'fidelity':
-        attention.NUM_ZERO = 8
-        attention.ORTHO = False
-        attention.ORTHO_v2 = True
-    elif mode == 'extremely style':
-        attention.NUM_ZERO = 16
-        attention.ORTHO = True
-        attention.ORTHO_v2 = False
-    else:
-        raise ValueError
+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 = 0.01
+
+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 = pipeline.get_id_embedding(id_image)
-        for supp_id_image in supp_images:
-            if supp_id_image is not None:
-                supp_id_image = resize_numpy_image_long(supp_id_image, 1024)
-                supp_id_embeddings = pipeline.get_id_embedding(supp_id_image)
-                id_embeddings = torch.cat(
-                    (id_embeddings, supp_id_embeddings if id_mix else supp_id_embeddings[:, :5]), dim=1
-                )
+        id_embeddings, uncond_id_embeddings = flux_generator.pulid_model.get_id_embedding(id_image, cal_uncond=use_true_cfg)
     else:
         id_embeddings = None
-
-    seed_everything(seed)
-    ims = []
-    for _ in range(n_samples):
-        img = pipeline.inference(prompt, (1, H, W), neg_prompt, id_embeddings, id_scale, scale, steps)[0]
-        ims.append(np.array(img))
-
-    return ims, pipeline.debug_img_list
-
+        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_ = '''
-<h2><b>Official 🤗 Gradio Demo</b></h2><h2><a href='https://github.com/ToTheBeginning/PuLID' target='_blank'><b>PuLID: Pure and Lightning ID Customization via Contrastive Alignment</b></a></h2>
+<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>
 
-**PuLID** is a tuning-free ID customization approach. PuLID maintains high ID fidelity while effectively reducing interference with the original model’s behavior.
+🚩 Updates:
+- 2024.11.01: update PuLID-FLUX-v0.9.1, please refer to <a href='https://github.com/ToTheBeginning/PuLID?tab=readme-ov-file#triangular_flag_on_post-updates'>our github repo</a> for more details.
 
-Code: <a href='https://github.com/ToTheBeginning/PuLID' target='_blank'>GitHub</a>. Techenical report: <a href='https://arxiv.org/abs/2404.16022' target='_blank'>ArXiv</a>.
 
 ❗️❗️❗️**Tips:**
-- we provide some examples in the bottom, you can try these example prompts first
-- a single ID image is usually sufficient, you can also supplement with additional auxiliary images
-- We offer two modes: fidelity mode and extremely style mode. In most cases, the default fidelity mode should suffice. If you find that the generated results are not stylized enough, you can choose the extremely style mode.
 
+- `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! [![GitHub Stars](https://img.shields.io/github/stars/ToTheBeginning/PuLID?style=social)](https://github.com/ToTheBeginning/PuLID)
+If PuLID is helpful, please help to ⭐ the <a href='https://github.com/ToTheBeginning/PuLID' target='_blank'> Github Repo</a>. Thanks!
 ---
-🚀 **Share**
-If you have generated satisfying or interesting images with PuLID, please share them with us or your friends!
-
-📝 **Citation**
-If you find our work useful for your research or applications, please cite using this bibtex:
-```bibtex
-@article{guo2024pulid,
-  title={PuLID: Pure and Lightning ID Customization via Contrastive Alignment},
-  author={Guo, Zinan and Wu, Yanze and Chen, Zhuowei and Chen, Lang and He, Qian},
-  journal={arXiv preprint arXiv:2404.16022},
-  year={2024}
-}
-```
-
-📋 **License**
-Apache-2.0 LICENSE. Please refer to the [LICENSE file](placeholder) for details.
-
 📧 **Contact**
-If you have any questions, feel free to open a discussion or contact us at <b>[email protected]</b> or <b>[email protected]</b>.
+If you have any questions or feedbacks, feel free to open a discussion or contact <b>[email protected]</b>.
 """  # noqa E501
 
-
-with gr.Blocks(title="PuLID", css=".gr-box {border-color: #8136e2}") as demo:
-    gr.Markdown(_HEADER_)
-    with gr.Row():
-        with gr.Column():
-            with gr.Row():
-                face_image = gr.Image(label="ID image (main)", sources="upload", type="numpy", height=256)
-                supp_image1 = gr.Image(
-                    label="Additional ID image (auxiliary)", sources="upload", type="numpy", height=256
-                )
-                supp_image2 = gr.Image(
-                    label="Additional ID image (auxiliary)", sources="upload", type="numpy", height=256
-                )
-                supp_image3 = gr.Image(
-                    label="Additional ID image (auxiliary)", sources="upload", type="numpy", height=256
-                )
-            prompt = gr.Textbox(label="Prompt", value='portrait,cinematic,wolf ears,white hair')
-            submit = gr.Button("Generate")
-            neg_prompt = gr.Textbox(label="Negative Prompt", value=DEFAULT_NEGATIVE_PROMPT)
-            scale = gr.Slider(
-                label="CFG, recommend value range [1, 1.5], 1 will be faster ",
-                value=1.2,
-                minimum=1,
-                maximum=1.5,
-                step=0.1,
-            )
-            n_samples = gr.Slider(label="Num samples", value=4, minimum=1, maximum=4, step=1)
-            seed = gr.Slider(
-                label="Seed", value=42, minimum=np.iinfo(np.uint32).min, maximum=np.iinfo(np.uint32).max, step=1
-            )
-            steps = gr.Slider(label="Steps", value=4, minimum=1, maximum=8, step=1)
-            with gr.Row():
-                H = gr.Slider(label="Height", value=1024, minimum=512, maximum=1280, step=64)
-                W = gr.Slider(label="Width", value=768, minimum=512, maximum=1280, step=64)
-            with gr.Row():
-                id_scale = gr.Slider(label="ID scale", minimum=0, maximum=5, step=0.05, value=0.8, interactive=True)
-                mode = gr.Dropdown(label="mode", choices=['fidelity', 'extremely style'], value='fidelity')
-                id_mix = gr.Checkbox(
-                    label="ID Mix (if you want to mix two ID image, please turn this on, otherwise, turn this off)",
-                    value=False,
-                )
-
-            gr.Markdown("## Examples")
-            example_inps = [
-                [
-                    'portrait,cinematic,wolf ears,white hair',
-                    'example_inputs/liuyifei.png',
-                    'fidelity',
-                ]
-            ]
-            gr.Examples(examples=example_inps, inputs=[prompt, face_image, mode], label='realistic')
-
-            example_inps = [
-                [
-                    'portrait, impressionist painting, loose brushwork, vibrant color, light and shadow play',
-                    'example_inputs/zcy.webp',
-                    'fidelity',
-                ]
-            ]
-            gr.Examples(examples=example_inps, inputs=[prompt, face_image, mode], label='painting style')
-
-            example_inps = [
-                [
-                    'portrait, flat papercut style, silhouette, clean cuts, paper, sharp edges, minimalist,color block,man',
-                    'example_inputs/lecun.jpg',
-                    'fidelity',
-                ]
-            ]
-            gr.Examples(examples=example_inps, inputs=[prompt, face_image, mode], label='papercut style')
-
-            example_inps = [
-                [
-                    'woman,cartoon,solo,Popmart Blind Box, Super Mario, 3d',
-                    'example_inputs/rihanna.webp',
-                    'fidelity',
-                ]
-            ]
-            gr.Examples(examples=example_inps, inputs=[prompt, face_image, mode], label='3d style')
-
-            example_inps = [
-                [
-                    'portrait, the legend of zelda, anime',
-                    'example_inputs/liuyifei.png',
-                    'extremely style',
+_DEV_DES = '''
+* Please refer to our repo for instructions on running gradio demo [locally](https://github.com/ToTheBeginning/PuLID/blob/main/docs/pulid_for_flux.md#local-gradio-demo)
+'''
+
+
+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:
+        with gr.Accordion("For Developers", open=False):
+            gr.Markdown(_DEV_DES)
+            
+        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, face_image, mode], label='anime style')
-
-            example_inps = [
-                [
-                    'portrait, superman',
-                    'example_inputs/lecun.jpg',
-                    'example_inputs/lifeifei.jpg',
-                    'fidelity',
-                    True,
+                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, face_image, supp_image1, mode, id_mix], label='id mix')
-
-        with gr.Column():
-            output = gr.Gallery(label='Output', elem_id="gallery")
-            intermediate_output = gr.Gallery(label='DebugImage', elem_id="gallery", visible=False)
-            gr.Markdown(_CITE_)
-
-    inps = [
-        face_image,
-        supp_image1,
-        supp_image2,
-        supp_image3,
-        prompt,
-        neg_prompt,
-        scale,
-        n_samples,
-        seed,
-        steps,
-        H,
-        W,
-        id_scale,
-        mode,
-        id_mix,
-    ]
-    submit.click(fn=run, inputs=inps, outputs=[output, intermediate_output])
-
-
-demo.launch()
+                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()