app.py

import os
import random
from datetime import datetime

import gradio as gr
import numpy as np
import torch
from diffusers import AutoencoderKL, DDIMScheduler
from einops import repeat
from huggingface_hub import hf_hub_download, snapshot_download
from omegaconf import OmegaConf
from PIL import Image
from torchvision import transforms
from transformers import CLIPVisionModelWithProjection

from src.models.pose_guider import PoseGuider
from src.models.unet_2d_condition import UNet2DConditionModel
from src.models.unet_3d import UNet3DConditionModel
from src.pipelines.pipeline_pose2vid_long import Pose2VideoPipeline
from src.utils.download_models import prepare_base_model, prepare_image_encoder
from src.utils.util import get_fps, read_frames, save_videos_grid

# Partial download
prepare_base_model()
prepare_image_encoder()

snapshot_download(
    repo_id="stabilityai/sd-vae-ft-mse", local_dir="./pretrained_weights/sd-vae-ft-mse"
)
snapshot_download(
    repo_id="patrolli/AnimateAnyone",
    local_dir="./pretrained_weights",
)


class AnimateController:
    def __init__(
        self,
        config_path="./configs/prompts/animation.yaml",
        weight_dtype=torch.float16,
    ):
        # Read pretrained weights path from config
        self.config = OmegaConf.load(config_path)
        self.pipeline = None
        self.weight_dtype = weight_dtype

    def animate(
        self,
        ref_image,
        pose_video_path,
        width=512,
        height=768,
        length=24,
        num_inference_steps=25,
        cfg=3.5,
        seed=123,
    ):
        generator = torch.manual_seed(seed)
        if isinstance(ref_image, np.ndarray):
            ref_image = Image.fromarray(ref_image)
        if self.pipeline is None:
            vae = AutoencoderKL.from_pretrained(
                self.config.pretrained_vae_path,
            ).to("cuda", dtype=self.weight_dtype)

            reference_unet = UNet2DConditionModel.from_pretrained(
                self.config.pretrained_base_model_path,
                subfolder="unet",
            ).to(dtype=self.weight_dtype, device="cuda")

            inference_config_path = self.config.inference_config
            infer_config = OmegaConf.load(inference_config_path)
            denoising_unet = UNet3DConditionModel.from_pretrained_2d(
                self.config.pretrained_base_model_path,
                self.config.motion_module_path,
                subfolder="unet",
                unet_additional_kwargs=infer_config.unet_additional_kwargs,
            ).to(dtype=self.weight_dtype, device="cuda")

            pose_guider = PoseGuider(320, block_out_channels=(16, 32, 96, 256)).to(
                dtype=self.weight_dtype, device="cuda"
            )

            image_enc = CLIPVisionModelWithProjection.from_pretrained(
                self.config.image_encoder_path
            ).to(dtype=self.weight_dtype, device="cuda")
            sched_kwargs = OmegaConf.to_container(infer_config.noise_scheduler_kwargs)
            scheduler = DDIMScheduler(**sched_kwargs)

            # load pretrained weights
            denoising_unet.load_state_dict(
                torch.load(self.config.denoising_unet_path, map_location="cpu"),
                strict=False,
            )
            reference_unet.load_state_dict(
                torch.load(self.config.reference_unet_path, map_location="cpu"),
            )
            pose_guider.load_state_dict(
                torch.load(self.config.pose_guider_path, map_location="cpu"),
            )

            pipe = Pose2VideoPipeline(
                vae=vae,
                image_encoder=image_enc,
                reference_unet=reference_unet,
                denoising_unet=denoising_unet,
                pose_guider=pose_guider,
                scheduler=scheduler,
            )
            pipe = pipe.to("cuda", dtype=self.weight_dtype)
            self.pipeline = pipe

        pose_images = read_frames(pose_video_path)
        src_fps = get_fps(pose_video_path)

        pose_list = []
        pose_tensor_list = []
        pose_transform = transforms.Compose(
            [transforms.Resize((height, width)), transforms.ToTensor()]
        )
        for pose_image_pil in pose_images[:length]:
            pose_list.append(pose_image_pil)
            pose_tensor_list.append(pose_transform(pose_image_pil))

        video = self.pipeline(
            ref_image,
            pose_list,
            width=width,
            height=height,
            video_length=length,
            num_inference_steps=num_inference_steps,
            guidance_scale=cfg,
            generator=generator,
        ).videos

        ref_image_tensor = pose_transform(ref_image)  # (c, h, w)
        ref_image_tensor = ref_image_tensor.unsqueeze(1).unsqueeze(0)  # (1, c, 1, h, w)
        ref_image_tensor = repeat(
            ref_image_tensor, "b c f h w -> b c (repeat f) h w", repeat=length
        )
        pose_tensor = torch.stack(pose_tensor_list, dim=0)  # (f, c, h, w)
        pose_tensor = pose_tensor.transpose(0, 1)
        pose_tensor = pose_tensor.unsqueeze(0)
        video = torch.cat([ref_image_tensor, pose_tensor, video], dim=0)

        save_dir = f"./output/gradio"
        if not os.path.exists(save_dir):
            os.makedirs(save_dir, exist_ok=True)
        date_str = datetime.now().strftime("%Y%m%d")
        time_str = datetime.now().strftime("%H%M")
        out_path = os.path.join(save_dir, f"{date_str}T{time_str}.mp4")
        save_videos_grid(
            video,
            out_path,
            n_rows=3,
            fps=src_fps,
        )

        torch.cuda.empty_cache()

        return out_path


controller = AnimateController()


def ui():
    with gr.Blocks() as demo:
        gr.HTML(
            """
            <h1 style="color:#dc5b1c;text-align:center">
                Moore-AnimateAnyone Gradio Demo 
            </h1>
            <div style="text-align:center">
            <div style="display: inline-block; text-align: left;">
            <p> This is a quick preview demo of Moore-AnimateAnyone. We appreciate the assistance provided by the HuggingFace team in setting up this demo. </p> 

            <p> To reduce waiting time, we limit the size (width, height and length) and inference steps when generating videos. </p>

            <p> If you like this project,  please consider giving a star on <a herf="https://github.com/MooreThreads/Moore-AnimateAnyone"> our GitHub repo </a> 🤗. </p>
            </div>
            </div>
            """
        )
        animation = gr.Video(
            format="mp4",
            label="Animation Results",
            height=448,
            autoplay=True,
        )

        with gr.Row():
            reference_image = gr.Image(label="Reference Image")
            motion_sequence = gr.Video(
                format="mp4", label="Motion Sequence", height=512
            )

            with gr.Column():
                width_slider = gr.Slider(
                    label="Width", minimum=256, maximum=448, value=448, step=64
                )
                height_slider = gr.Slider(
                    label="Height", minimum=256, maximum=512, value=512, step=64
                )
                length_slider = gr.Slider(
                    label="Video Length", minimum=24, maximum=24, value=24, step=1
                )
                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: gr.Textbox.update(value=random.randint(1, 1e8)),
                        inputs=[],
                        outputs=[seed_textbox],
                    )
                with gr.Row():
                    sampling_steps = gr.Slider(
                        label="Sampling steps",
                        value=15,
                        info="default: 15",
                        step=5,
                        maximum=15,
                        minimum=10,
                    )
                    guidance_scale = gr.Slider(
                        label="Guidance scale",
                        value=3.5,
                        info="default: 3.5",
                        step=0.5,
                        maximum=6.5,
                        minimum=2.0,
                    )
                submit = gr.Button("Animate")

        def read_video(video):
            return video

        def read_image(image):
            return Image.fromarray(image)

        # when user uploads a new video
        motion_sequence.upload(
            read_video, motion_sequence, motion_sequence, queue=False
        )
        # when `first_frame` is updated
        reference_image.upload(
            read_image, reference_image, reference_image, queue=False
        )
        # when the `submit` button is clicked
        submit.click(
            controller.animate,
            [
                reference_image,
                motion_sequence,
                width_slider,
                height_slider,
                length_slider,
                sampling_steps,
                guidance_scale,
                seed_textbox,
            ],
            animation,
        )

        # Examples
        gr.Markdown("## Examples")
        gr.Examples(
            examples=[
                [
                    "./configs/inference/ref_images/anyone-5.png",
                    "./configs/inference/pose_videos/anyone-video-2_kps.mp4",
                ],
                [
                    "./configs/inference/ref_images/anyone-10.png",
                    "./configs/inference/pose_videos/anyone-video-1_kps.mp4",
                ],
                [
                    "./configs/inference/ref_images/anyone-2.png",
                    "./configs/inference/pose_videos/anyone-video-5_kps.mp4",
                ],
            ],
            inputs=[reference_image, motion_sequence],
            outputs=animation,
        )

    return demo


demo = ui()
demo.queue(max_size=10)
demo.launch(share=True, show_api=False)