如何基于火山引擎弹性容器快速部署 MagicAnimate 应用

冯明星2023年12月14日约 8 分钟

火山引擎弹性容器实例（VCI）是一种 Serverless 和容器化的计算服务，旨在帮助企业控制云成本、专注于构建应用本身。

来源 | 火山引擎云原生团队

近日，字节跳动和新加坡国立大学联合推出了一款“人体图像动画生成工具”MagicAnimate，它能够根据用户指定的人物图片和动作序列，生成对应的动效视频。简单来说，用户只需提供一张人物照片，再配合骨骼动画引导，就能生成自然的动画视频，如下图所示：

根据相关论文 MagicAnimate: Temporally Consistent Human Image Animation using Diffusion Model 介绍，MagicAnimate 在 TikTok 舞蹈数据集上，生成的视频保真度比最强基线提高了超过 38%。

目前，MagicAnimate 已在 GitHub 正式开源（地址：github.com/magic-research/magic-animate），为了方便更多企业用户体验，本文将按以下步骤，介绍如何通过火山引擎弹性容器实例 VCI 快速拉起 MagicAnimate 应用，部署属于你的 AI 动画生成工具。

构建容器镜像
制作 VCI 容器镜像缓存
通过 VCI 部署 magic-animate 应用
浏览器请求测试效果

什么是 VCI

这一年以来，AIGC 模型在技术和成果上的突破令业界激动无比，但是对于模型开发者和相关企业来说，模型的商业部署应用仍然是首要问题。

火山引擎弹性容器实例（VCI）是一种 Serverless 和容器化的计算服务，它旨在帮助企业控制云成本、专注于构建应用本身——无需购买和管理底层云服务器等基础设施，仅为容器实际运行消耗的资源付费。

面对 AIGC 应用部署场景，VCI 能快速拉起实例，在流量高峰时快速扩充新资源，在流量低谷时进行缩容，保障企业资源完全按需使用，天然适合帮助 AIGC 类应用应对潮汐流量和突发流量请求。

构建容器镜像

首先，我们需要下载模型：

python3from huggingface_hub import snapshot_downloadsnapshot_download(repo_id="runwayml/stable-diffusion-v1-5",local_dir="/root/magic-animate/sdv15")snapshot_download(repo_id="stabilityai/sd-vae-ft-mse",local_dir="/root/magic-animate/vae")snapshot_download(repo_id="zcxu-eric/MagicAnimate",local_dir="/root/magic-animate/MagicAnimate")

将相关模型放到项目的 pretrained_models 目录下，目录结构 Dockerfile 如下所示（此处省略具体镜像制作过程），同时将模型也打包到容器镜像中。

可用镜像地址：paas-cn-beijing.cr.volces.com/aigc/magic-animate:v1

FROM paas-cn-beijing.cr.volces.com/cuda/cuda:11.4.3-devel-ubuntu20.04-torch
LABEL org.opencontainers.image.authors="[email protected]"RUN apt-get update && apt-get install -y git  RUN git clone https://github.com/magic-research/magic-animate.git && cd magic-animate && pip3 install -r requirements.txtCOPY demo/gradio_animate.py magic-animate/demo/RUN mkdir -p magic-animate/pretrained_models COPY pretrained_models magic-animate/pretrained_modelsWORKDIR magic-animateCMD ["python3","-m","demo.gradio_animate"]

需要修改的demo/gradio_animate.py如下，制作容器镜像时会将该文件覆盖项目原有的相同名称的文件，达到修改监听地址的目的。

# Copyright 2023 ByteDance and/or its affiliates.## Copyright (2023) MagicAnimate Authors## ByteDance, its affiliates and licensors retain all intellectual# property and proprietary rights in and to this material, related# documentation and any modifications thereto. Any use, reproduction,# disclosure or distribution of this material and related documentation# without an express license agreement from ByteDance or# its affiliates is strictly prohibited.import argparseimport imageioimport numpy as npimport gradio as grfrom PIL import Image
from demo.animate import MagicAnimate
animator = MagicAnimate()
def animate(reference_image, motion_sequence_state, seed, steps, guidance_scale):    return animator(reference_image, motion_sequence_state, seed, steps, guidance_scale)
with gr.Blocks() as demo:
    gr.HTML(        """        <div style="display: flex; justify-content: center; align-items: center; text-align: center;">        <a href="https://github.com/magic-research/magic-animate" style="margin-right: 20px; text-decoration: none; display: flex; align-items: center;">        </a>        <div>            <h1 >MagicAnimate: Temporally Consistent Human Image Animation using Diffusion Model</h1>            <h5 style="margin: 0;">If you like our project, please give us a star ✨ on Github for the latest update.</h5>            <div style="display: flex; justify-content: center; align-items: center; text-align: center;>                <a href="https://arxiv.org/abs/2311.16498"><img src="https://img.shields.io/badge/Arxiv-2311.16498-red"></a>                <a href='https://showlab.github.io/magicanimate'><img src='https://img.shields.io/badge/Project_Page-MagicAnimate-green' alt='Project Page'></a>                <a href='https://github.com/magic-research/magic-animate'><img src='https://img.shields.io/badge/Github-Code-blue'></a>            </div>        </div>        </div>        """)    animation = gr.Video(format="mp4", label="Animation Results", autoplay=True)    with gr.Row():        reference_image  = gr.Image(label="Reference Image")        motion_sequence  = gr.Video(format="mp4", label="Motion Sequence")        with gr.Column():            random_seed         = gr.Textbox(label="Random seed", value=1, info="default: -1")            sampling_steps      = gr.Textbox(label="Sampling steps", value=25, info="default: 25")            guidance_scale      = gr.Textbox(label="Guidance scale", value=7.5, info="default: 7.5")            submit              = gr.Button("Animate")
    def read_video(video):        reader = imageio.get_reader(video)        fps = reader.get_meta_data()['fps']        return video    def read_image(image, size=512):        return np.array(Image.fromarray(image).resize((size, size)))    # when user uploads a new video    motion_sequence.upload(        read_video,        motion_sequence,        motion_sequence    )    # when `first_frame` is updated    reference_image.upload(        read_image,        reference_image,        reference_image    )    # when the `submit` button is clicked    submit.click(        animate,        [reference_image, motion_sequence, random_seed, sampling_steps, guidance_scale],         animation    )
    # Examples    gr.Markdown("## Examples")    gr.Examples(        examples=[            ["inputs/applications/source_image/monalisa.png", "inputs/applications/driving/densepose/running.mp4"],            ["inputs/applications/source_image/demo4.png", "inputs/applications/driving/densepose/demo4.mp4"],            ["inputs/applications/source_image/0002.png", "inputs/applications/driving/densepose/demo4.mp4"],            ["inputs/applications/source_image/dalle2.jpeg", "inputs/applications/driving/densepose/running2.mp4"],            ["inputs/applications/source_image/dalle8.jpeg", "inputs/applications/driving/densepose/dancing2.mp4"],            ["inputs/applications/source_image/multi1_source.png", "inputs/applications/driving/densepose/multi_dancing.mp4"],        ],        inputs=[reference_image, motion_sequence],        outputs=animation,    )

demo.launch(server_name='0.0.0.0')#修改监听地址，默认为127.0.0.1

制作 VCI 容器镜像缓存

与传统应用相比，AIGC 类应用具有容器镜像大的明显特征，通常在 10GB 以上。在不加速的情况下，10 GB 以上容器镜像的拉取用时较久，再加上在弹性场景，拉取镜像是非常频繁的操作，这会大幅延长 GPU 资源不可用时间，造成 GPU 资源浪费。

火山引擎提供镜像加速方案，可以将镜像拉取用时从分钟级降低到秒级，帮助用户大幅提升效率（详见《AIGC 推理加速：火山引擎镜像加速实践》）。这里我们直接使用加速效果最好的镜像缓存方案。

通过火山引擎控制台，我们可以非常快速地制作容器镜像缓存，极大提升镜像拉取速度。

通过 VCI 部署 magic-animate 应用

使用 YAML 方式部署

apiVersion: apps/v1kind: Deploymentmetadata:  name: magic-animate  namespace: defaultspec:  progressDeadlineSeconds: 600  replicas: 1  revisionHistoryLimit: 10  selector:    matchLabels:      app: magic-animate  strategy:    rollingUpdate:      maxSurge: 25%      maxUnavailable: 25%    type: RollingUpdate  template:    metadata:      annotations:        vci.vke.volcengine.com/preferred-instance-types: vci.ini2.26c-243gi        vke.volcengine.com/burst-to-vci: enforce      creationTimestamp: null      labels:        app: magic-animate    spec:      containers:      - image: paas-cn-beijing.cr.volces.com/aigc/magic-animate:v1        imagePullPolicy: IfNotPresent        name: app        resources:          limits:            nvidia.com/gpu: "1"        terminationMessagePath: /dev/termination-log        terminationMessagePolicy: File      dnsPolicy: ClusterFirst      nodeSelector:        kubernetes.io/hostname: vci-node1-cn-beijing-b      restartPolicy: Always      schedulerName: default-scheduler      securityContext: {}      terminationGracePeriodSeconds: 30

创建服务供请求访问

apiVersion: v1kind: Servicemetadata:  annotations:    service.beta.kubernetes.io/volcengine-loadbalancer-address-type: PUBLIC    service.beta.kubernetes.io/volcengine-loadbalancer-bandwidth: "10"    service.beta.kubernetes.io/volcengine-loadbalancer-eip-billing-type: "3"    service.beta.kubernetes.io/volcengine-loadbalancer-health-check-flag: "off"    service.beta.kubernetes.io/volcengine-loadbalancer-ip-version: ipv4    service.beta.kubernetes.io/volcengine-loadbalancer-isp-type: BGP    service.beta.kubernetes.io/volcengine-loadbalancer-name: magic-animate-lb    service.beta.kubernetes.io/volcengine-loadbalancer-pass-through: "true"    service.beta.kubernetes.io/volcengine-loadbalancer-scheduler: wrr    service.beta.kubernetes.io/volcengine-loadbalancer-spec: small_1    service.beta.kubernetes.io/volcengine-loadbalancer-subnet-id: subnet-mjn0iia7j6yo5smt1a9fcz5v  name: magic-animate-svc  namespace: defaultspec:  allocateLoadBalancerNodePorts: true  externalTrafficPolicy: Local  internalTrafficPolicy: Cluster  ipFamilies:  - IPv4  ipFamilyPolicy: SingleStack  ports:  - port: 80    protocol: TCP    targetPort: 7860  selector:    app: magic-animate  sessionAffinity: None  type: LoadBalancer