每周学点 AI：ComfyUI + Modal 的一键部署脚本

AntBlack2025-09-07 17:50

一. 前言

又在研究薅羊毛的方案 ，还是在 Modal 平台上面进行的处理 ，这一期直接把 comfyUI 部署上去了，再也不担心没积分跑图了。

这一篇就带大家好好的薅羊毛 ，我们是专业的。

二. 平台介绍

@ juejin.cn/post/754425...

这里上一篇我们介绍过这个平台 ，有的时候会有一些活动 ，总的来说每个月有30$ 的金额 ，算下来平均生图100张以上应该是没问题 ，多注册几个号也是一个不错的途径。

整体界面

  • 部署完会有2个组件 ，可以看到里面的对应的部署的 File 文件体系

外挂卷 Volumns

  • 一般情况下 ，由于模型的体积较大 ，就需要把模型丢到 Volumes 里面 ，避免每次都需要进行下载
  • 官方提供的文档里面支持直接从 HuggingFace 里面下载 ，我这里 扩展了其他的加载渠道

三. 上代码

python 复制代码 
# =============================================================================
# ComfyUI Flux API 服务
#
# S1: 环境准备 - 导入依赖包和构建基础镜像
# S2: 模型下载 - 从HuggingFace和远程URL下载模型文件
# S3: 服务配置 - 创建Modal应用和存储卷
# S4: UI服务 - 提供交互式Web界面
# S5: API服务 - 提供图像生成API接口
# =============================================================================

# --- S1: 环境准备阶段 ---
import json
import os
import subprocess
import uuid
from pathlib import Path
from typing import Dict

import modal
import modal.experimental
import requests

# S1.1: 构建基础Docker镜像，安装Python环境和基础依赖
image = (
    modal.Image.debian_slim(python_version="3.11")
    .apt_install("git")
    .pip_install("fastapi[standard]==0.115.4")
    .pip_install("comfy-cli==1.4.1")
    .pip_install("requests==2.32.3")
    .run_commands("comfy --skip-prompt install --fast-deps --nvidia --version 0.3.41")
)

# S1.2: 安装ComfyUI自定义节点扩展
image = image.run_commands(
    "comfy node install --fast-deps was-node-suite-comfyui@1.0.2")

# --- S2: 模型下载阶段 ---


def hf_download():
    """
    S2: 下载所需的AI模型文件
    S2.1: 从HuggingFace下载基础模型和LoRA模型
    S2.2: 从远程URL下载额外的模型文件
    S2.3: 创建软链接到ComfyUI模型目录
    """
    from huggingface_hub import hf_hub_download

    # S2.1: 下载HuggingFace基础模型
    print("📥 S2.1: 开始下载Flux基础模型...")
    flux_model = hf_hub_download(
        repo_id="Comfy-Org/flux1-dev",
        filename="flux1-dev-fp8.safetensors",
        cache_dir="/cache",
    )
    subprocess.run(
        f"ln -s {flux_model} /root/comfy/ComfyUI/models/checkpoints/flux1-dev-fp8.safetensors",
        shell=True,
        check=True,
    )

    # S2.2: 下载HuggingFace LoRA模型列表
    print("📥 S2.2: 开始下载LoRA模型...")
    lora_dir = "/root/comfy/ComfyUI/models/loras"
    os.makedirs(lora_dir, exist_ok=True)

    lora_models = [
        {
            "repo_id": "UmeAiRT/FLUX.1-dev-LoRA-Ume_Sky",
            "filename": "ume_sky_v2.safetensors",
            "local_name": "ume_sky_v2.safetensors"
        },
        {
            "repo_id": "Shakker-Labs/FLUX.1-dev-LoRA-Dark-Fantasy",
            "filename": "FLUX.1-dev-lora-Dark-Fantasy.safetensors",
            "local_name": "FLUX.1-dev-lora-Dark-Fantasy.safetensors"
        },
        {
            "repo_id": "aleksa-codes/flux-ghibsky-illustration",
            "filename": "lora_v2.safetensors",
            "local_name": "lora_v2.safetensors"
        }
    ]

    for lora in lora_models:
        try:
            print(f"  📦 下载LoRA: {lora['repo_id']}")
            lora_path = hf_hub_download(
                repo_id=lora["repo_id"],
                filename=lora["filename"],
                cache_dir="/cache",
            )
            subprocess.run(
                f"ln -s {lora_path} /root/comfy/ComfyUI/models/loras/{lora['local_name']}",
                shell=True,
                check=True
            )
        except Exception as e:
            print(f"❌ LoRA下载失败 {lora['repo_id']}: {e}")

    # S2.3: 从远程URL下载额外模型
    print("📥 S2.3: 开始从远程URL下载模型...")
    url_models = [
        {
            "url": "https://civitai-delivery-worker-prod.5ac0637cfd0766c97916cefa3764fbdf.r2.cloudflarestorage.com/model/9230585/1752195403.4GHo.safetensors?X-Amz-Expires=86400&response-content-disposition=attachment%3B%20filename%3D%221752195403.safetensors%22&X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=e01358d793ad6966166af8b3064953ad/20250907/us-east-1/s3/aws4_request&X-Amz-Date=20250907T062143Z&X-Amz-SignedHeaders=host&X-Amz-Signature=3814daa1e9d0c91d1992b32341f64a5d27dcd3135babb7975c1b36914d893617",
            "filename": "1752195403.safetensors",
            "type": "loras"
        },
        {
            "url": "https://civitai-delivery-worker-prod.5ac0637cfd0766c97916cefa3764fbdf.r2.cloudflarestorage.com/model/17651/flux1LoraFlywayEpic.NKkZ.safetensors?X-Amz-Expires=86400&response-content-disposition=attachment%3B%20filename%3D%22flux.1_lora_flyway_Epic-Characters_v1.safetensors%22&X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=e01358d793ad6966166af8b3064953ad/20250907/us-east-1/s3/aws4_request&X-Amz-Date=20250907T083947Z&X-Amz-SignedHeaders=host&X-Amz-Signature=4ae9dcbd8c0205fb258b7839bb5895a94db6831ab9bf87c10936f4eafd6c028a",
            "filename": "中世纪风格.safetensors",
            "type": "loras"
        },
        {
            "url": "https://civitai-delivery-worker-prod.5ac0637cfd0766c97916cefa3764fbdf.r2.cloudflarestorage.com/model/933225/newFantasyCorev4FLUX.pt13.safetensors?X-Amz-Expires=86400&response-content-disposition=attachment%3B%20filename%3D%22New_Fantasy_CoreV4_FLUX.safetensors%22&X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=e01358d793ad6966166af8b3064953ad/20250907/us-east-1/s3/aws4_request&X-Amz-Date=20250907T084003Z&X-Amz-SignedHeaders=host&X-Amz-Signature=45b8f8e990b9105872964a0d6a440b131bcdfcde0e0d8d0d5de29756b24b55d9",
            "filename": "幻想风.safetensors",
            "type": "loras"
        },
        {
            "url": "https://civitai-delivery-worker-prod.5ac0637cfd0766c97916cefa3764fbdf.r2.cloudflarestorage.com/model/4768839/fluxthous40k.YPhQ.safetensors?X-Amz-Expires=86400&response-content-disposition=attachment%3B%20filename%3D%22FluxThouS40k.safetensors%22&X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=e01358d793ad6966166af8b3064953ad/20250907/us-east-1/s3/aws4_request&X-Amz-Date=20250907T084019Z&X-Amz-SignedHeaders=host&X-Amz-Signature=2c3aa9cd675bd52d2a190ffc033b6f98bf434d990b1053be3da5885e57571aa5",
            "filename": "铠甲风.safetensors",
            "type": "loras"
        },
        {
            "url": "https://civitai-delivery-worker-prod.5ac0637cfd0766c97916cefa3764fbdf.r2.cloudflarestorage.com/model/65967/neonfantasyprimeflux.uMmv.safetensors?X-Amz-Expires=86400&response-content-disposition=attachment%3B%20filename%3D%22NeonFantasyPrimeFLUX-000049.safetensors%22&X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=e01358d793ad6966166af8b3064953ad/20250907/us-east-1/s3/aws4_request&X-Amz-Date=20250907T084033Z&X-Amz-SignedHeaders=host&X-Amz-Signature=1a8b596f5218a1f6a56b808907f7bbf72213383b5b35a3a580d731f232810537",
            "filename": "异世界.safetensors",
            "type": "loras"
        }
    ]

    url_download_dir = "/cache"
    os.makedirs(url_download_dir, exist_ok=True)

    for model in url_models:
        final_model_path = os.path.join(
            "/root/comfy/ComfyUI/models", model["type"], model["filename"])

        # 检查文件是否已存在
        if os.path.exists(final_model_path):
            print(f"✅ 模型 '{model['filename']}' 已存在，跳过下载")
            continue

        print(f"⬇️ 正在下载 '{model['filename']}' ...")
        cached_file_path = os.path.join(url_download_dir, model["filename"])

        try:
            with requests.get(model["url"], stream=True, allow_redirects=True) as r:
                r.raise_for_status()
                with open(cached_file_path, 'wb') as f:
                    for chunk in r.iter_content(chunk_size=8192):
                        f.write(chunk)

            # 创建软链接到ComfyUI目标目录
            os.makedirs(os.path.dirname(final_model_path), exist_ok=True)
            subprocess.run(
                f"ln -s {cached_file_path} {final_model_path}", shell=True, check=True)
            print(f"✅ 模型 '{model['filename']}' 下载并链接完成")

        except Exception as e:
            print(f"❌ URL下载失败 {model['url']}: {e}")


# --- S3: 服务配置阶段 ---
print("🔧 S3: 开始配置Modal服务...")

# S3.1: 创建持久化存储卷
vol = modal.Volume.from_name("hf-hub-cache", create_if_missing=True)

# S3.2: 完成镜像构建，添加HuggingFace支持和模型文件
image = (
    image.pip_install("huggingface_hub[hf_transfer]==0.34.4")
    .env({"HF_HUB_ENABLE_HF_TRANSFER": "1"})
    .run_function(hf_download, volumes={"/cache": vol})
    .add_local_file(
        Path(__file__).parent / "workflow_api.json",
        "/root/workflow_api.json"
    )
    .add_local_file(
        local_path="C:\\Users\\zzg\Downloads\\F.1风格模型 _ Miboso光影纹理大师_V1.safetensors",
        remote_path="/root/comfy/ComfyUI/models/loras/F.1风格模型 _ Miboso光影纹理大师_V1.safetensors"
    )
)

# S3.3: 创建Modal应用实例
app = modal.App(name="example-comfyapp", image=image)

# --- S4: UI服务阶段 ---


@app.function(max_containers=1, gpu="L40S", volumes={"/cache": vol})
@modal.concurrent(max_inputs=10)
@modal.web_server(8000, startup_timeout=60)
def ui():
    """
    S4: 提供ComfyUI交互式Web界面服务
    - 启动ComfyUI Web服务器
    - 监听0.0.0.0:8000端口
    - 支持最多10个并发用户
    """
    print("🌐 S4: 启动ComfyUI交互式Web界面...")
    subprocess.Popen(
        "comfy launch -- --listen 0.0.0.0 --port 8000",
        shell=True
    )

# --- S5: API服务阶段 ---


@app.cls(scaledown_window=300, gpu="L40S", volumes={"/cache": vol})
@modal.concurrent(max_inputs=5)
class ComfyUI:
    """
    S5: ComfyUI API服务类
    提供图像生成的RESTful API接口
    支持最多5个并发请求处理
    """
    port: int = 8000

    @modal.enter()
    def launch_comfy_background(self):
        """
        S5.1: 容器启动时初始化ComfyUI后台服务
        """
        print(f"🚀 S5.1: 启动ComfyUI后台服务，端口: {self.port}")
        cmd = f"comfy launch --background -- --port {self.port}"
        subprocess.run(cmd, shell=True, check=True)

    @modal.method()
    def infer(self, workflow_path: str = "/root/workflow_api.json"):
        """
        S5.2: 执行图像生成推理
        - 检查服务健康状态
        - 运行ComfyUI工作流
        - 返回生成的图像字节数据
        """
        print("🎨 S5.2: 开始执行图像生成推理...")

        # S5.2.1: 检查服务健康状态
        self.poll_server_health()

        # S5.2.2: 执行工作流
        cmd = f"comfy run --workflow {workflow_path} --wait --timeout 1200 --verbose"
        subprocess.run(cmd, shell=True, check=True)

        # S5.2.3: 获取生成的图像文件
        output_dir = "/root/comfy/ComfyUI/output"
        workflow = json.loads(Path(workflow_path).read_text())
        file_prefix = [
            node.get("inputs")
            for node in workflow.values()
            if node.get("class_type") == "SaveImage"
        ][0]["filename_prefix"]

        # S5.2.4: 返回图像字节数据
        for f in Path(output_dir).iterdir():
            if f.name.startswith(file_prefix):
                return f.read_bytes()

    @modal.fastapi_endpoint(method="POST")
    def api(self, item: Dict):
        """
        S5.3: FastAPI端点 - 处理HTTP POST请求
        - 接收用户提示词
        - 生成唯一的工作流文件
        - 调用推理方法生成图像
        - 返回图像响应
        """
        from fastapi import Response

        print("📡 S5.3: 处理API请求...")

        # S5.3.1: 加载工作流模板
        workflow_data = json.loads(
            (Path(__file__).parent / "workflow_api.json").read_text()
        )

        # S5.3.2: 设置用户提示词
        workflow_data["6"]["inputs"]["text"] = item["prompt"]

        # S5.3.3: 生成唯一的客户端ID和文件名
        client_id = uuid.uuid4().hex
        workflow_data["9"]["inputs"]["filename_prefix"] = client_id

        # S5.3.4: 保存自定义工作流文件
        new_workflow_file = f"{client_id}.json"
        json.dump(workflow_data, Path(new_workflow_file).open("w"))

        # S5.3.5: 执行推理并返回图像
        img_bytes = self.infer.local(new_workflow_file)
        return Response(img_bytes, media_type="image/jpeg")

    def poll_server_health(self) -> Dict:
        """
        S5.4: 健康检查 - 确保ComfyUI服务正常运行
        """
        import socket
        import urllib

        try:
            req = urllib.request.Request(
                f"http://127.0.0.1:{self.port}/system_stats"
            )
            urllib.request.urlopen(req, timeout=5)
            print("✅ ComfyUI服务健康检查通过")
        except (socket.timeout, urllib.error.URLError) as e:
            print("❌ ComfyUI服务健康检查失败，停止容器")
            modal.experimental.stop_fetching_inputs()
            raise Exception(
                "ComfyUI server is not healthy, stopping container")

四. 代码执行

  • modal deploy comfyapp_multi.py
shell 复制代码 
D:\code\python\demo\AI_DEMO\modal_comfyui\comfyapp_multi.py:125: SyntaxWarning: invalid escape sequence '\D'
  local_path="C:\\Users\\xxx\Downloads\\F.1风格模型 _ Miboso光影纹理大师_V1.safetensors",
Building image im-CPvMHYnP5JcYjFTX5iDN0d

=> Step 0: running function 'hf_download'
⬇️ Downloading '1752195403.safetensors' from URL...
✅ Download and link for '1752195403.safetensors' complete.
Saving image...
Image saved, took 1.10s
Finished image build for im-CPvMHYnP5JcYjFTX5iDN0d
✓ Created objects.
├── 🔨 Created mount D:\code\python\demo\AI_DEMO\modal_comfyui\comfyapp_multi.py
├── 🔨 Created mount C:\Users\xx\Downloads\F.1风格模型 _ Miboso光影纹理大师_V1.safetensors
├── 🔨 Created mount D:\code\python\demo\AI_DEMO\modal_comfyui\workflow_api.json
├── 🔨 Created function hf_download.
├── 🔨 Created web function ui => https://xxx--example-comfyapp-ui.modal.run
├── 🔨 Created function ComfyUI.*.
└── 🔨 Created web endpoint for ComfyUI.api => https://xxx--example-comfyapp-comfyui-api.modal.run
✓ App deployed in 30.436s! 🎉

View Deployment: https://modal.com/apps/xxx/main/deployed/example-comfyapp

五. 扩展功能

同步的还撸了一个 Client 端 ，通过 API 调用 ，但是内部没有控制参数 ，有需求的也可以用 Client ：

总结

@ gitee.com/antblack/ai...

消耗还好 ，基本上够用 ，功能也和常规的平台不大 ，可搞。

最后的最后 ❤️❤️❤️👇👇👇

相关推荐
IT古董12 分钟前
【第五章:计算机视觉-项目实战之图像分割实战】1.图像分割理论-(2)图像分割衍生：语义分割、实例分割、弱监督语义分割
人工智能·计算机视觉
猿究院-陆昱泽16 分钟前
Redis 五大核心数据结构知识点梳理
redis·后端·中间件
yuriy.wang43 分钟前
Spring IOC源码篇五 核心方法obtainFreshBeanFactory.doLoadBeanDefinitions
java·后端·spring
大明者省2 小时前
《青花》歌曲,使用3D表现出意境
人工智能
一朵小红花HH2 小时前
SimpleBEV:改进的激光雷达-摄像头融合架构用于三维目标检测
论文阅读·人工智能·深度学习·目标检测·机器学习·计算机视觉·3d
Daitu_Adam2 小时前
R语言——ggmap包可视化地图
人工智能·数据分析·r语言·数据可视化
weixin_377634842 小时前
【阿里DeepResearch】写作组件WebWeaver详解
人工智能
AndrewHZ2 小时前
【AI算力系统设计分析】1000PetaOps 算力云计算系统设计方案（大模型训练推理专项版）
人工智能·深度学习·llm·云计算·模型部署·大模型推理·算力平台
咖啡教室3 小时前
程序员应该掌握的网络命令telnet、ping和curl
运维·后端
AI_gurubar3 小时前
[NeurIPS‘25] AI infra / ML sys 论文（解析）合集
人工智能
热门推荐
01GitHub 镜像站点02UV 工具安装与国内镜像源配置指南03Multisim使用教程详尽版--(2025最新版)04KGG转MP3工具|非KGM文件|解密音频05UV安装并设置国内源0646个Nano-banana 精选提示词，持续更新中07Spec-Kit 使用指南08VsCode远程Copilot无法使用Claude Agent问题09jdk21下载、安装（Windows、Linux、macOS）10阿里最新开源Wan2.2-Animate-14B 本地部署教程：统一双模态框架，MoE架构赋能电影级角色动画与替换