ComfyUI进阶实战：用工作流实现AI电商批量自动化生图

# 克隆 ComfyUI
git clone https://github.com/comfyanonymous/ComfyUI.git

# 进入目录
cd ComfyUI

# 安装依赖（需要有Python 3.10+和CUDA环境）
pip install -r requirements.txt

# 启动
python main.py

# 在 ComfyUI 根目录执行
cd custom_nodes
git clone https://github.com/ltdrdata/ComfyUI-Manager.git

# 在 ComfyUI/models 目录下
# 下载Flux-dev（推荐）
# 模型约23GB

# 或者Flux-schnell（更小更快）
# 模型约12GB

# 下载Text Encoder
mkdir -p models/clip
# 从HuggingFace下载flux-text-encoders
# T5XXL: t5xxl_fp8_e4m3fn.safetensors
# CLIP-L: clip_l.safetensors

# IP-Adapter模型：ip-adapter_sd15.bin（适合Flux）
# 模式：Progressive Rescale（渐进式缩放，保持产品结构）
# 权重：0.6-0.8（太低一致性差，太高场景受限）
# 噪声注入：0.3（增加场景多样性）

"high-end fashion editorial photography, [产品类型],
[场景描述：如modern minimalist studio backdrop],
natural soft lighting from left window,
clean aesthetic, professional fashion photography,
white background transition to [场景], 
product remains centered and unaltered,
8k, cinematic, fashion magazine quality"

"fashion editorial, women's cotton summer dress,
white dress displayed in luxury boutique setting,
warm golden hour lighting, natural shadows,
professional fashion photography, 
clean white background seamlessly blending to 
minimalist indoor setting, product-centric,
8k, Vogue magazine quality, realistic fabric texture"

"fashion editorial, women's cotton summer dress,
modern coffee shop scene, natural window light,
model wearing the dress, lifestyle photography,
professional fashion photography, 
warm tones, product clearly visible,
8k, editorial quality, realistic fabric drape"

# ImagePadForFlow Node
# 在场景图上留出文字区域

# TextOverlay Node
# 在指定位置叠加指定文字
# 支持字体选择（中文字体需额外安装）
# 支持字体大小、颜色、描边等设置

# InstantID节点参数
face_kps: [参考图关键点]
model: antelopev2（推荐，准确度高）
strength: 0.7-0.85

# 第1帧（建立角色基准）
"portrait photo of [角色描述],
close-up, neutral expression,
professional photography, studio lighting,
realistic, high detail face,
consistent with reference image"
附加：上传角色设定图作为参考

# 第2-N帧（同一角色，不同场景/动作）
"same character as reference image,
[具体动作描述：sitting at desk, looking surprised],
[场景描述],
maintain exact facial features and identity from reference,
consistent character design across all frames,
anime/manga style, high quality"
附加：继续使用同一张参考图

python main.py --listen 0.0.0.0 --port 8188 --enable-cors

import requests
import json
import base64

# 1. 提交生图任务
def queue_prompt(prompt_dict):
    """提交生图任务到ComfyUI"""
    response = requests.post(
        "http://localhost:8188/prompt",
        json={"prompt": prompt_dict}
    )
    return response.json()

# 2. 查询任务状态
def get_history(prompt_id):
    """获取任务结果"""
    response = requests.get(
        f"http://localhost:8188/history/{prompt_id}"
    )
    return response.json()

# 3. 下载生成结果
def get_image(filename):
    """下载生成的图片"""
    response = requests.get(
        f"http://localhost:8188/view?filename={filename}"
    )
    return response.content

import requests
import json
import os
import time
from pathlib import Path

class EcommerceBatchGenerator:
    def __init__(self, api_url="http://localhost:8188"):
        self.api = api_url
        self.workflows = self._load_workflows()
    
    def _load_workflows(self):
        """加载预设工作流JSON"""
        with open("workflows/product_scene.json", "r") as f:
            return json.load(f)
    
    def batch_generate_product_scenes(self, product_folder, output_folder, 
                                        scene_prompts):
        """
        批量生成产品场景图
        
        Args:
            product_folder: 产品白底图文件夹路径
            output_folder: 输出文件夹路径
            scene_prompts: 场景提示词列表
        """
        os.makedirs(output_folder, exist_ok=True)
        
        product_files = list(Path(product_folder).glob("*.png"))
        
        for product_file in product_files:
            for i, prompt in enumerate(scene_prompts):
                # 加载工作流模板
                workflow = self._prepare_workflow(
                    self.workflows["product_scene"],
                    product_file,
                    prompt
                )
                
                # 提交任务
                result = self.queue_prompt(workflow)
                prompt_id = result["prompt_id"]
                
                # 等待完成（轮询）
                output_file = None
                while not output_file:
                    time.sleep(3)
                    history = self.get_history(prompt_id)
                    if prompt_id in history:
                        outputs = history[prompt_id].get("outputs", {})
                        for node_id, node_data in outputs.items():
                            if "images" in node_data:
                                output_file = node_data["images"][0]["filename"]
                
                # 保存结果
                image_data = self.get_image(output_file)
                output_name = f"{product_file.stem}_scene{i+1}.png"
                with open(os.path.join(output_folder, output_name), "wb") as f:
                    f.write(image_data)
                
                print(f"✅ {product_file.name} - Scene {i+1} 完成")
    
    def _prepare_workflow(self, template, product_path, scene_prompt):
        """替换工作流参数"""
        workflow = template.copy()
        # 替换产品图片路径
        workflow["load_image"]["inputs"]["image_path"] = str(product_path)
        # 替换场景提示词
        workflow["clip_text_encode"]["inputs"]["text"] = scene_prompt
        return workflow
    
    def queue_prompt(self, prompt_dict):
        return requests.post(f"{self.api}/prompt", json={"prompt": prompt_dict}).json()
    
    def get_history(self, prompt_id):
        return requests.get(f"{self.api}/history/{prompt_id}").json()
    
    def get_image(self, filename):
        return requests.get(f"{self.api}/view?filename={filename}").content

# 使用示例
generator = EcommerceBatchGenerator()

scene_prompts = [
    "modern minimalist living room, warm natural light, product centered",
    "office desk setup, product displayed prominently, professional photography",
    "outdoor garden setting, soft daylight, lifestyle shot",
    "luxury boutique display, elegant lighting, premium atmosphere"
]

generator.batch_generate_product_scenes(
    product_folder="product_images/whitebg",
    output_folder="output/scenes",
    scene_prompts=scene_prompts
)

# Schnell模型显存需求约8GB，比dev版低40%
# 速度更快，适合批量生成
# 质量略低，但电商用图足够

# FP8量化版模型，约节省30%显存
# 从HuggingFace下载flux1-dev-fp8.tar

# 把大图片拆成小块处理，再拼合
# 适合超高分辨率主图（4000×4000px+）

# 使用Turbo采样器替代默认Euler
# 速度提升2-3倍
sampler_name: "euler_ancestral"  # 推荐电商使用
steps: 20  # Flux-schnell仅需4-8步

# 同时生成多张图，而非逐张生成
# 显存允许的情况下，batch_size设为2-4
# 效率提升显著

# ComfyUI模型目录可能很大，用软链接共享模型
# 适用于多用户或工作室环境

# Windows
mklink /D "C:\ComfyUI\models\unet" "D:\SharedModels\flux"

# Linux/Mac
ln -s /mnt/shared/models/flux ~/ComfyUI/models/unet/flux