app-huggingface.py

import json
import os
import os.path as osp
import random
from argparse import ArgumentParser
from datetime import datetime

import gradio as gr
import numpy as np
import torch
from huggingface_hub import hf_hub_download
from omegaconf import OmegaConf
from PIL import Image

from animatediff.pipelines import I2VPipeline
from animatediff.utils.util import RANGE_LIST, save_videos_grid

sample_idx = 0

css = """
.toolbutton {
    margin-buttom: 0em 0em 0em 0em;
    max-width: 2.5em;
    min-width: 2.5em !important;
    height: 2.5em;
}
"""

parser = ArgumentParser()
parser.add_argument('--config', type=str, default='example/config/base.yaml')
parser.add_argument('--server-name', type=str, default='0.0.0.0')
parser.add_argument('--port', type=int, default=7860)
parser.add_argument('--share', action='store_true')
parser.add_argument('--local-debug', action='store_true')

parser.add_argument('--save-path', default='samples')

args = parser.parse_args()
LOCAL_DEBUG = args.local_debug


BASE_CONFIG = 'example/config/base.yaml'
STYLE_CONFIG_LIST = {
    '3d_cartoon': './example/openxlab/3-3d.yaml',
    'realistic': './example/openxlab/1-realistic.yaml',
}


# download models
PIA_PATH = './models/PIA'
VAE_PATH = './models/VAE'
DreamBooth_LoRA_PATH = './models/DreamBooth_LoRA'


if not LOCAL_DEBUG:
    CACHE_PATH = './models'

    PIA_PATH = osp.join(CACHE_PATH, 'PIA')
    VAE_PATH = osp.join(CACHE_PATH, 'VAE')
    DreamBooth_LoRA_PATH = osp.join(CACHE_PATH, 'DreamBooth_LoRA')
    STABLE_DIFFUSION_PATH = osp.join(CACHE_PATH, 'StableDiffusion')

    os.makedirs(PIA_PATH, exist_ok=True)
    os.makedirs(VAE_PATH, exist_ok=True)
    os.makedirs(DreamBooth_LoRA_PATH, exist_ok=True)
    os.makedirs(STABLE_DIFFUSION_PATH, exist_ok=True)

    PIA_PATH = hf_hub_download(repo_id='Leoxing/PIA',
                               filename='pia.ckpt', cache_dir=PIA_PATH)
    PIA_PATH = '/'.join(PIA_PATH.split('/')[:-1])
    os.system('bash download_bashscripts/1-RealisticVision.sh')
    os.system('bash download_bashscripts/2-RcnzCartoon.sh')
    # hf_hub_download(repo_id='frankjoshua/realisticVisionV51_v51VAE',
    #                 filename='realisticVisionV51_v51VAE.safetensors',
    #                 cache_dir='')
    print(os.listdir(DreamBooth_LoRA_PATH))

    # unet
    unet_full_path = hf_hub_download(repo_id='runwayml/stable-diffusion-v1-5',
                                     subfolder='unet', filename='diffusion_pytorch_model.bin',
                                     cache_dir='models/StableDiffusion')
    STABLE_DIFFUSION_PATH = '/'.join(unet_full_path.split('/')[:-2])
    hf_hub_download(repo_id='runwayml/stable-diffusion-v1-5',
                    subfolder='unet', filename='config.json',
                    cache_dir='models/StableDiffusion')

    # vae
    hf_hub_download(repo_id='runwayml/stable-diffusion-v1-5',
                    subfolder='vae', filename='config.json',
                    cache_dir='models/StableDiffusion')
    hf_hub_download(repo_id='runwayml/stable-diffusion-v1-5',
                    subfolder='vae', filename='diffusion_pytorch_model.bin',
                    cache_dir='models/StableDiffusion')

    # text encoder
    hf_hub_download(repo_id='runwayml/stable-diffusion-v1-5',
                    subfolder='text_encoder', filename='config.json',
                    cache_dir='models/StableDiffusion')
    hf_hub_download(repo_id='runwayml/stable-diffusion-v1-5',
                    subfolder='text_encoder', filename='pytorch_model.bin',
                    cache_dir='models/StableDiffusion')

    # tokenizer
    hf_hub_download(repo_id='runwayml/stable-diffusion-v1-5',
                    subfolder='tokenizer', filename='merges.txt',
                    cache_dir='models/StableDiffusion')
    hf_hub_download(repo_id='runwayml/stable-diffusion-v1-5',
                    subfolder='tokenizer', filename='special_tokens_map.json',
                    cache_dir='models/StableDiffusion')
    hf_hub_download(repo_id='runwayml/stable-diffusion-v1-5',
                    subfolder='tokenizer', filename='tokenizer_config.json',
                    cache_dir='models/StableDiffusion')
    hf_hub_download(repo_id='runwayml/stable-diffusion-v1-5',
                    subfolder='tokenizer', filename='vocab.json',
                    cache_dir='models/StableDiffusion')

    # scheduler
    hf_hub_download(repo_id='runwayml/stable-diffusion-v1-5',
                    subfolder='scheduler', filename='scheduler_config.json',
                    cache_dir='models/StableDiffusion')

    # model index
    hf_hub_download(repo_id='runwayml/stable-diffusion-v1-5', filename='model_index.json',
                    cache_dir='models/StableDiffusion')

else:
    PIA_PATH = './models/PIA'
    VAE_PATH = './models/VAE'
    DreamBooth_LoRA_PATH = './models/DreamBooth_LoRA'
    STABLE_DIFFUSION_PATH = './models/StableDiffusion/sd15'


def preprocess_img(img_np, max_size: int = 512):

    ori_image = Image.fromarray(img_np).convert('RGB')

    width, height = ori_image.size

    short_edge = max(width, height)
    if short_edge > max_size:
        scale_factor = max_size / short_edge
    else:
        scale_factor = 1
    width = int(width * scale_factor)
    height = int(height * scale_factor)
    ori_image = ori_image.resize((width, height))

    if (width % 8 != 0) or (height % 8 != 0):
        in_width = (width // 8) * 8
        in_height = (height // 8) * 8
    else:
        in_width = width
        in_height = height
        in_image = ori_image

    in_image = ori_image.resize((in_width, in_height))
    in_image_np = np.array(in_image)
    return in_image_np, in_height, in_width


class AnimateController:
    def __init__(self):

        # config dirs
        self.basedir = os.getcwd()
        self.savedir = os.path.join(
            self.basedir, args.save_path, datetime.now().strftime("Gradio-%Y-%m-%dT%H-%M-%S"))
        self.savedir_sample = os.path.join(self.savedir, "sample")
        os.makedirs(self.savedir, exist_ok=True)

        self.inference_config = OmegaConf.load(args.config)
        self.style_configs = {k: OmegaConf.load(
            v) for k, v in STYLE_CONFIG_LIST.items()}

        self.pipeline_dict = self.load_model_list()

    def load_model_list(self):
        print(os.listdir(DreamBooth_LoRA_PATH))
        pipeline_dict = dict()
        for style, cfg in self.style_configs.items():
            dreambooth_path = cfg.get('dreambooth', 'none')
            if dreambooth_path and dreambooth_path.upper() != 'NONE':
                dreambooth_path = osp.join(
                    DreamBooth_LoRA_PATH, dreambooth_path)
            lora_path = cfg.get('lora', None)
            if lora_path is not None:
                lora_path = osp.join(DreamBooth_LoRA_PATH, lora_path)
            lora_alpha = cfg.get('lora_alpha', 0.0)
            vae_path = cfg.get('vae', None)
            if vae_path is not None:
                vae_path = osp.join(VAE_PATH, vae_path)

            pipeline_dict[style] = I2VPipeline.build_pipeline(
                self.inference_config,
                STABLE_DIFFUSION_PATH,
                unet_path=osp.join(PIA_PATH, 'pia.ckpt'),
                dreambooth_path=dreambooth_path,
                lora_path=lora_path,
                lora_alpha=lora_alpha,
                vae_path=vae_path,
                ip_adapter_path='h94/IP-Adapter',
                ip_adapter_scale=0.1)
        return pipeline_dict

    def fetch_default_n_prompt(self, style: str):
        cfg = self.style_configs[style]
        n_prompt = cfg.get('n_prompt', '')
        ip_adapter_scale = cfg.get('real_ip_adapter_scale', 0)

        gr.Info('Set default negative prompt and ip_adapter_scale.')
        print('Set default negative prompt and ip_adapter_scale.')

        return n_prompt, ip_adapter_scale

    def animate(
        self,
        init_img,
        motion_scale,
        prompt_textbox,
        negative_prompt_textbox,
        sample_step_slider,
        cfg_scale_slider,
        seed_textbox,
        ip_adapter_scale,
        style,
        progress=gr.Progress(),
    ):

        if seed_textbox != -1 and seed_textbox != "":
            torch.manual_seed(int(seed_textbox))
        else:
            torch.seed()
        seed = torch.initial_seed()

        pipeline = self.pipeline_dict[style]
        init_img, h, w = preprocess_img(init_img)
        sample = pipeline(
            image=init_img,
            prompt=prompt_textbox,
            negative_prompt=negative_prompt_textbox,
            num_inference_steps=sample_step_slider,
            guidance_scale=cfg_scale_slider,
            width=w,
            height=h,
            video_length=16,
            mask_sim_template_idx=motion_scale - 1,
            ip_adapter_scale=ip_adapter_scale,
            progress_fn=progress,
        ).videos

        save_sample_path = os.path.join(
            self.savedir_sample, f"{sample_idx}.mp4")
        save_videos_grid(sample, save_sample_path)

        sample_config = {
            "prompt": prompt_textbox,
            "n_prompt": negative_prompt_textbox,
            "num_inference_steps": sample_step_slider,
            "guidance_scale": cfg_scale_slider,
            "width": w,
            "height": h,
            "seed": seed,
            "motion": motion_scale,
        }
        json_str = json.dumps(sample_config, indent=4)
        with open(os.path.join(self.savedir, "logs.json"), "a") as f:
            f.write(json_str)
            f.write("\n\n")

        return save_sample_path


controller = AnimateController()


def ui():
    with gr.Blocks(css=css) as demo:

        gr.HTML(
            "<div align='center'><font size='7'> <img src=\"file/pia.png\" style=\"height: 72px;\"/ > Your Personalized Image Animator</font></div>"
            "<div align='center'><font size='7'>via Plug-and-Play Modules in Text-to-Image Models </font></div>"
        )
        with gr.Row():
            gr.Markdown(
                "<div align='center'><font size='5'><a href='https://pi-animator.github.io/'>Project Page</a> &ensp;"  # noqa
                "<a href='https://arxiv.org/abs/2312.13964/'>Paper</a> &ensp;"
                "<a href='https://github.com/open-mmlab/PIA'>Code</a> &ensp;"  # noqa
                "Try More Style: <a href='https://openxlab.org.cn/apps/detail/zhangyiming/PiaPia-AnimationStyle'>Click here! </a></font></div>"  # noqa
            )

        with gr.Row(equal_height=False):
            with gr.Column():
                with gr.Row():
                    init_img = gr.Image(label='Input Image')

                style_dropdown = gr.Dropdown(label='Style', choices=list(
                    STYLE_CONFIG_LIST.keys()), value=list(STYLE_CONFIG_LIST.keys())[0])

                with gr.Row():
                    prompt_textbox = gr.Textbox(label="Prompt", lines=1)
                    gift_button = gr.Button(
                        value='🎁', elem_classes='toolbutton'
                    )

                def append_gift(prompt):
                    rand = random.randint(0, 2)
                    if rand == 1:
                        prompt = prompt + 'wearing santa hats'
                    elif rand == 2:
                        prompt = prompt + 'lift a Christmas gift'
                    else:
                        prompt = prompt + 'in Christmas suit, lift a Christmas gift'
                    gr.Info('Merry Christmas! Add magic to your prompt!')
                    return prompt

                gift_button.click(
                    fn=append_gift,
                    inputs=[prompt_textbox],
                    outputs=[prompt_textbox],
                )

                motion_scale_silder = gr.Slider(
                    label='Motion Scale (Larger value means larger motion but less identity consistency)',
                    value=1, step=1, minimum=1, maximum=len(RANGE_LIST))
                ip_adapter_scale = gr.Slider(
                    label='IP-Apdater Scale', value=controller.fetch_default_n_prompt(
                        list(STYLE_CONFIG_LIST.keys())[0])[1], minimum=0, maximum=1)

                with gr.Accordion('Advance Options', open=False):
                    negative_prompt_textbox = gr.Textbox(
                        value=controller.fetch_default_n_prompt(
                            list(STYLE_CONFIG_LIST.keys())[0])[0],
                        label="Negative prompt", lines=2)

                    sample_step_slider = gr.Slider(
                        label="Sampling steps", value=20, minimum=10, maximum=100, step=1)

                    cfg_scale_slider = gr.Slider(
                        label="CFG Scale", value=7.5, minimum=0, maximum=20)

                    with gr.Row():
                        seed_textbox = gr.Textbox(label="Seed", value=-1)
                        seed_button = gr.Button(
                            value="\U0001F3B2", elem_classes="toolbutton")
                    seed_button.click(
                        fn=lambda x: random.randint(1, 1e8),
                        outputs=[seed_textbox],
                        queue=False
                    )

                generate_button = gr.Button(
                    value="Generate", variant='primary')

            result_video = gr.Video(
                label="Generated Animation", interactive=False)

        style_dropdown.change(fn=controller.fetch_default_n_prompt,
                              inputs=[style_dropdown],
                              outputs=[negative_prompt_textbox,
                                       ip_adapter_scale],
                              queue=False)

        generate_button.click(
            fn=controller.animate,
            inputs=[
                init_img,
                motion_scale_silder,
                prompt_textbox,
                negative_prompt_textbox,
                sample_step_slider,
                cfg_scale_slider,
                seed_textbox,
                ip_adapter_scale,
                style_dropdown,
            ],
            outputs=[result_video]
        )

        def create_example(input_list):
            return gr.Examples(
                examples=input_list,
                inputs=[
                    init_img,
                    result_video,
                    prompt_textbox,
                    negative_prompt_textbox,
                    style_dropdown,
                    motion_scale_silder,
                ],
            )

        gr.Markdown(
            '### Merry Christmas!'
        )
        create_example(
            [
                [
                    './__assets__/image_animation/yiming/yiming.jpeg',
                    './__assets__/image_animation/yiming/yiming.mp4',
                    '1boy in Christmas suit, lift a Christmas gift',
                    'wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg',
                    '3d_cartoon',
                    2,
                ],
                [
                    './__assets__/image_animation/yanhong/yanhong.png',
                    './__assets__/image_animation/yanhong/yanhong.mp4',
                    '1girl lift a Christmas gift',
                    'wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg',
                    '3d_cartoon',
                    2,
                ],
            ],

        )

        with gr.Accordion('More Examples for Style Transfer', open=False):
            create_example([
                [

                    './__assets__/image_animation/style_transfer/anya/anya.jpg',
                    './__assets__/image_animation/style_transfer/anya/2.mp4',
                    '1girl open mouth ',
                    'wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg',
                    '3d_cartoon',
                    3,
                ],
                [
                    './__assets__/image_animation/magnitude/genshin/genshin.jpg',
                    './__assets__/image_animation/magnitude/genshin/3.mp4',
                    'cherry blossoms in the wind, raidenshogundef, yaemikodef, best quality, 4k',
                    'wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg',
                    '3d_cartoon',
                    3,
                ],

            ])

        with gr.Accordion('More Examples for Prompt Changing', open=False):
            create_example(
                [
                    [
                        './__assets__/image_animation/real/lighthouse.jpg',
                        './__assets__/image_animation/real/1.mp4',
                        'lightning, lighthouse',
                        'wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg',
                        'realistic',
                        1,
                    ],
                    [
                        './__assets__/image_animation/real/lighthouse.jpg',
                        './__assets__/image_animation/real/2.mp4',
                        'sun rising, lighthouse',
                        'wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg',
                        'realistic',
                        1,
                    ],
                    [
                        './__assets__/image_animation/real/lighthouse.jpg',
                        './__assets__/image_animation/real/3.mp4',
                        'fireworks, lighthouse',
                        'wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg',
                        'realistic',
                        1,
                    ],
                    [
                        './__assets__/image_animation/rcnz/harry.png',
                        './__assets__/image_animation/rcnz/1.mp4',
                        '1boy smiling',
                        'wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg',
                        '3d_cartoon',
                        2
                    ],
                    [
                        './__assets__/image_animation/rcnz/harry.png',
                        './__assets__/image_animation/rcnz/2.mp4',
                        '1boy playing magic fire',
                        'wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg',
                        '3d_cartoon',
                        2
                    ],
                    [
                        './__assets__/image_animation/rcnz/harry.png',
                        './__assets__/image_animation/rcnz/3.mp4',
                        '1boy is waving hands',
                        'wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg',
                        '3d_cartoon',
                        2
                    ]
                ])

            with gr.Accordion('Examples for Motion Magnitude', open=False):
                create_example(
                    [
                        [
                            './__assets__/image_animation/magnitude/labrador.png',
                            './__assets__/image_animation/magnitude/1.mp4',
                            'cherry blossoms in the wind, raidenshogundef, yaemikodef, best quality, 4k',
                            'wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg',
                            '3d_cartoon',
                            1,
                        ],
                        [
                            './__assets__/image_animation/magnitude/labrador.png',
                            './__assets__/image_animation/magnitude/2.mp4',
                            'cherry blossoms in the wind, raidenshogundef, yaemikodef, best quality, 4k',
                            'wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg',
                            '3d_cartoon',
                            2,
                        ],
                        [
                            './__assets__/image_animation/magnitude/labrador.png',
                            './__assets__/image_animation/magnitude/3.mp4',
                            'cherry blossoms in the wind, raidenshogundef, yaemikodef, best quality, 4k',
                            'wrong white balance, dark, sketches,worst quality,low quality, deformed, distorted, disfigured, bad eyes, wrong lips,weird mouth, bad teeth, mutated hands and fingers, bad anatomy,wrong anatomy, amputation, extra limb, missing limb, floating,limbs, disconnected limbs, mutation, ugly, disgusting, bad_pictures, negative_hand-neg',
                            '3d_cartoon',
                            3,
                        ]
                    ])

    return demo


if __name__ == "__main__":
    demo = ui()
    demo.queue(max_size=10)
    demo.launch(server_name=args.server_name,
                server_port=args.port, share=args.share,
                max_threads=40,
                allowed_paths=['pia.png'])