B9. AI 产品图片与视频生成 Pipeline

路径: Path B: 技术人 · 模块: B9 最后更新: 2026-07-31 难度: 高级 预计时间: 每天 1 小时，2-3 周 前置模块: 无（独立模块，但建议了解 A7 视觉内容）

章节导航

1. 为什么需要 AI 图片 Pipeline · 2. 技术栈选择 · 3. ComfyUI 产品图工作流 · 4. 云端 API 方案 · 5. 批量生成 Pipeline · 6. 视频生成 · 7. 质量控制与合规 · 8. 常见陷阱 · 9. 完成标志

本模块你将构建

• 一个 ComfyUI 产品图生成工作流（白底主图 + 场景图 + 信息图）
• 一个 API 驱动的批量图片生成 Pipeline（Midjourney/GPT Image 2/FLUX.2）
• 一个产品视频自动生成系统
• 品牌视觉一致性保障机制

核心理念：电商产品图是转化率的第一要素。传统方式是请摄影师拍摄（$500-2000/产品），AI 方式是用 ComfyUI/Midjourney 生成（$0-50/产品）。但 AI 生成不是“一键出图“，需要构建可重复、可控制、品牌一致的 Pipeline。

相关阅读: A7 视觉内容 运营视角的 AI 视觉内容方法论

1. 为什么需要 AI 图片 Pipeline

1.1 电商图片需求矩阵

1.2 AI 图片生成的挑战

2. 技术栈选择

2.1 方案对比

2.2 推荐组合

推荐的 AI 图片技术栈：

主图/场景图生成：
ComfyUI + Flux（本地，完全控制）
或 Midjourney（云端，质量最高）
或 GPT Image 2 API（可编程，批量生成）

后处理：
rembg（Python 去背景）
Pillow（图片处理、文字叠加）
OpenCV（高级图片处理）

批量管理：
Python 脚本（自动化工作流）
Canva Brand Kit（模板管理）

3. ComfyUI 产品图工作流

3.1 安装 ComfyUI

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

# 安装依赖
pip3 install -r requirements.txt

# 下载模型（Flux 推荐）
# 将模型文件放到 models/checkpoints/ 目录

# 启动
python3 main.py
# 浏览器打开 http://127.0.0.1:8188

3.2 产品图生成工作流

真实案例：ComfyUI 产品图工作流实战 MyAIForce 展示了一个完整的 ComfyUI 产品图工作流：输入一张护肤品图片和描述性 Prompt，工作流自动将产品无缝融入新背景，调整光照和阴影以匹配新环境，确保自然和谐的外观。工作流包含 7 个步骤：上传图片→设置背景→基础调整→产品定位→重新打光→重绘→恢复细节（MyAIForce）。

Content rephrased for compliance with licensing restrictions.

真实案例：Midjourney + ComfyUI 组合工作流 另一个高级工作流将 Midjourney 和 ComfyUI 结合：先用 Midjourney 生成高质量的场景背景，再用 ComfyUI 的 ControlNet 和 IP-Adapter 将产品精确放置到场景中，同时调整光照和阴影以保留产品文字等关键细节（MyAIForce）。

Content rephrased for compliance with licensing restrictions.

真实案例：ComfyUI 背景替换 V4 工作流 最新的 V4 背景替换工作流使用 SDXL checkpoints，仅需 10 个采样步骤和约 6GB VRAM 即可完成基础任务。使用 Flux 模型可以获得更高质量的效果，但需要更多 VRAM（MyAIForce）。

Content rephrased for compliance with licensing restrictions.

ComfyUI 电商产品图完整工作流（7 步）：

Step 1: 上传图片和设置背景
Load Image 节点：加载产品实拍图
背景选择：上传预设背景 或 用 Prompt 生成
参数设置：分辨率、采样步数

Step 2: 基础调整
产品抠图（Florence2Run 或 rembg）
尺寸调整
初始合成

Step 3: 产品定位
调整产品在画面中的位置
缩放比例
角度调整

Step 4: 重新打光（Relighting）
IC-Light 节点：根据背景调整产品光照
阴影方向匹配
高光调整

Step 5: 生成背景
Flux Fill + Redux：生成与产品匹配的背景
或 IP-Adapter：复制参考图片的风格
KSampler：执行生成

Step 6: 重绘（Inpainting）
修复产品与背景的接缝
添加自然阴影
细节融合

Step 7: 恢复细节和颜色
恢复产品原始颜色
锐化细节
最终输出
保存为 PNG/JPEG

3.3 电商场景 Prompt 模板（40+ 测试过的模板）

真实资源：Apatero 整理了 40+ 经过测试的 AI 产品图 Prompt 模板，覆盖白底、场景、平铺、信息图等所有电商场景（Apatero）。

Content rephrased for compliance with licensing restrictions.

# 电商产品图 Prompt 模板库（扩展版）
PROMPT_TEMPLATES = {
    # === 主图系列 ===
    "amazon_main": {
        "positive": "professional product photography, {product}, centered on pure white background #FFFFFF, product fills 85 percent of frame, studio lighting with soft shadows, high resolution 8k, sharp focus, no text no logos no watermarks, commercial catalog style",
        "negative": "blurry, low quality, text, watermark, logo, human, hand, colored background, shadow on background, props, accessories not part of product"
    },
    "shopify_hero": {
        "positive": "hero product shot, {product}, clean minimal background with subtle gradient, dramatic studio lighting, slight shadow underneath, premium feel, editorial quality, 4k",
        "negative": "cluttered, busy background, text, watermark, low quality"
    },

    # === 场景图系列 ===
    "lifestyle_home": {
        "positive": "lifestyle product photography, {product} in modern minimalist home, natural window lighting, warm tones, shallow depth of field, bokeh background, editorial style, authentic feel",
        "negative": "artificial, oversaturated, studio look, text, watermark"
    },
    "lifestyle_outdoor": {
        "positive": "outdoor lifestyle photography, {product} in natural setting, golden hour lighting, vibrant colors, adventure feel, authentic, editorial quality",
        "negative": "indoor, artificial lighting, text, watermark, studio"
    },
    "lifestyle_office": {
        "positive": "modern office setting, {product} on clean desk, natural lighting from window, minimalist decor, professional atmosphere, shallow depth of field",
        "negative": "cluttered, messy, dark, text, watermark"
    },
    "lifestyle_kitchen": {
        "positive": "modern kitchen setting, {product} on marble countertop, natural lighting, fresh ingredients nearby, clean and bright, food photography style",
        "negative": "dirty, cluttered, dark, text, watermark"
    },

    # === 平铺图系列 ===
    "flat_lay_minimal": {
        "positive": "flat lay photography, {product} with complementary items, top-down view, clean arrangement on {surface}, soft shadows, minimalist, {color_scheme}",
        "negative": "cluttered, messy, blurry, text, 3D perspective"
    },
    "flat_lay_seasonal": {
        "positive": "seasonal flat lay, {product} surrounded by {season} elements, top-down view, cohesive color palette, editorial styling, natural textures",
        "negative": "cluttered, artificial, text, watermark"
    },

    # === 信息图背景系列 ===
    "infographic_clean": {
        "positive": "clean infographic background for {product}, {color_scheme} gradient, modern design, ample negative space for text overlay, professional, soft lighting on product",
        "negative": "text, numbers, charts, cluttered, busy, distracting elements"
    },
    "infographic_comparison": {
        "positive": "split comparison layout background, {product} centered, left side and right side clearly divided, clean modern design, space for before/after or feature comparison text",
        "negative": "text, numbers, cluttered"
    },

    # === 社交媒体系列 ===
    "instagram_aesthetic": {
        "positive": "instagram aesthetic product shot, {product}, trendy styling, {color_scheme} color palette, natural lighting, lifestyle feel, square format, influencer style",
        "negative": "corporate, boring, text, watermark, low quality"
    },
    "tiktok_dynamic": {
        "positive": "dynamic product shot, {product}, vibrant colors, energetic composition, slight motion blur on background, youth-oriented, vertical format 9:16",
        "negative": "static, boring, corporate, text"
    },

    # === A+ Content 系列 ===
    "aplus_brand_story": {
        "positive": "brand story photography, {product} in aspirational setting, warm emotional lighting, lifestyle context, premium quality, cinematic feel",
        "negative": "cheap, low quality, text, watermark"
    },
    "aplus_feature_highlight": {
        "positive": "close-up detail shot, {product} {feature} highlighted, macro photography style, sharp focus on detail, soft background, studio lighting",
        "negative": "blurry, wide shot, text, watermark"
    }
}

def generate_prompt(template_name: str, product: str, **kwargs) -> dict:
    """生成产品图 Prompt"""
    template = PROMPT_TEMPLATES[template_name]
    # 填充默认值
    defaults = {
        "surface": "white marble",
        "color_scheme": "blue and white",
        "season": "autumn",
        "feature": "texture detail"
    }
    for k, v in defaults.items():
        kwargs.setdefault(k, v)

    return {
        "positive": template["positive"].format(product=product, **kwargs),
        "negative": template["negative"]
    }

# 使用示例
prompt = generate_prompt(
    "lifestyle_home",
    product="wireless bluetooth earbuds with charging case"
)
print(prompt["positive"])

4. 云端 API 方案

4.1 GPT Image 2 批量生成

from openai import OpenAI
import requests
from pathlib import Path

client = OpenAI()

def generate_product_image(
    product_description: str,
    style: str = "white_background",
    size: str = "1024x1024",
    output_dir: str = "output"
) -> str:
    """用 GPT Image 2 生成产品图"""

    prompts = {
        "white_background": f"Professional product photography of {product_description}, centered on pure white background, studio lighting, high resolution, commercial quality",
        "lifestyle": f"Lifestyle product photography of {product_description} being used in a modern home setting, natural lighting, warm tones, editorial quality",
        "amazon_main": f"Amazon product listing main image: {product_description}, pure white background (#FFFFFF), product fills 85% of frame, no text or logos, professional studio photography"
    }

    response = client.images.generate(
        model="gpt-image-2",
        prompt=prompts[style],
        size=size,
        quality="hd",
        n=1
    )

    # 下载图片
    image_url = response.data[0].url
    Path(output_dir).mkdir(exist_ok=True)

    img_data = requests.get(image_url).content
    filepath = f"{output_dir}/{product_description[:30]}_{style}.png"
    with open(filepath, "wb") as f:
        f.write(img_data)

    return filepath

# 批量生成
products = [
    "wireless bluetooth earbuds with charging case",
    "stainless steel water bottle 32oz",
    "portable neck fan with LED display"
]

for product in products:
    for style in ["white_background", "lifestyle"]:
        path = generate_product_image(product, style)
        print(f"Generated: {path}")

4.2 去背景 + 白底合成

from rembg import remove
from PIL import Image
import io

def create_amazon_main_image(input_path: str, output_path: str):
    """创建 Amazon 合规的白底主图"""
    # 读取图片
    with open(input_path, "rb") as f:
        input_data = f.read()

    # 去背景
    output_data = remove(input_data)

    # 创建白底画布
    fg = Image.open(io.BytesIO(output_data)).convert("RGBA")

    # 计算产品占比（Amazon 要求 85%+）
    bbox = fg.getbbox()
    product_w = bbox[2] - bbox[0]
    product_h = bbox[3] - bbox[1]

    # 创建正方形白底（产品占 85%）
    canvas_size = int(max(product_w, product_h) / 0.85)
    canvas = Image.new("RGBA", (canvas_size, canvas_size), (255, 255, 255, 255))

    # 居中放置产品
    offset_x = (canvas_size - product_w) // 2 - bbox[0]
    offset_y = (canvas_size - product_h) // 2 - bbox[1]
    canvas.paste(fg, (offset_x, offset_y), fg)

    # 保存为 RGB（Amazon 不接受透明背景）
    canvas.convert("RGB").save(output_path, "JPEG", quality=95)
    print(f"Amazon main image saved: {output_path}")

5. 批量生成 Pipeline

5.1 完整的产品图生成 Pipeline

import os
import json
from pathlib import Path
from datetime import datetime
from dataclasses import dataclass
from typing import Optional

@dataclass
class ProductImageRequest:
    """产品图生成请求"""
    product_name: str
    product_description: str
    source_image: Optional[str] = None # 产品实拍图路径
    brand_color: str = "blue"
    target_platforms: list = None # ["amazon", "shopify", "instagram"]

    def __post_init__(self):
        if self.target_platforms is None:
            self.target_platforms = ["amazon", "shopify"]

class ProductImagePipeline:
    """电商产品图批量生成 Pipeline"""

    def __init__(self, method: str = "openai", output_dir: str = "output/images"):
        self.method = method
        self.output_dir = output_dir
        Path(output_dir).mkdir(parents=True, exist_ok=True)
        self.log = []

    def generate_product_set(self, request: ProductImageRequest) -> dict:
        """为一个产品生成完整的图片集"""
        product_dir = os.path.join(
            self.output_dir,
            request.product_name.replace(" ", "_")[:30]
        )
        Path(product_dir).mkdir(exist_ok=True)

        results = {"product": request.product_name, "images": {}}

        # 1. Amazon 白底主图
        if "amazon" in request.target_platforms:
            self._log(f"生成 Amazon 主图: {request.product_name}")
            main_img = self._generate_image(
                request, "amazon_main",
                os.path.join(product_dir, "amazon_main.jpg")
            )
            # 后处理：去背景 + 白底合成
            amazon_img = self._post_process_amazon(main_img)
            results["images"]["amazon_main"] = amazon_img

            # 合规检查
            compliance = check_amazon_compliance(amazon_img)
            results["images"]["amazon_compliance"] = compliance
            if not compliance["compliant"]:
                self._log(f" Amazon 合规问题: {compliance['issues']}")

        # 2. 场景图 x3
        scenes = [
            ("modern living room", "lifestyle_home"),
            ("outdoor natural setting", "lifestyle_outdoor"),
            ("clean office desk", "lifestyle_office")
        ]
        results["images"]["lifestyle"] = []
        for i, (scene, template) in enumerate(scenes):
            self._log(f"生成场景图 {i+1}/3: {scene}")
            img = self._generate_image(
                request, template,
                os.path.join(product_dir, f"lifestyle_{i+1}.jpg"),
                scene=scene
            )
            results["images"]["lifestyle"].append(img)

        # 3. 信息图背景 x2
        results["images"]["infographic"] = []
        for i, color in enumerate(["blue and white", "warm earth tones"]):
            self._log(f"生成信息图背景 {i+1}/2")
            img = self._generate_image(
                request, "infographic_clean",
                os.path.join(product_dir, f"infographic_{i+1}.jpg"),
                color_scheme=color
            )
            results["images"]["infographic"].append(img)

        # 4. 社交媒体素材
        if "instagram" in request.target_platforms:
            self._log("生成 Instagram 素材")
            img = self._generate_image(
                request, "instagram_aesthetic",
                os.path.join(product_dir, "instagram.jpg"),
                color_scheme=request.brand_color
            )
            results["images"]["instagram"] = img

        # 5. A+ Content 品牌故事图
        self._log("生成 A+ Content 图")
        img = self._generate_image(
            request, "aplus_brand_story",
            os.path.join(product_dir, "aplus_brand.jpg")
        )
        results["images"]["aplus"] = img

        # 保存元数据
        metadata = {
            "product": request.product_name,
            "generated_at": datetime.now().isoformat(),
            "method": self.method,
            "images": {k: str(v) for k, v in results["images"].items()},
            "log": self.log
        }
        with open(os.path.join(product_dir, "metadata.json"), "w") as f:
            json.dump(metadata, f, indent=2, ensure_ascii=False)

        self._log(f" 完成: {request.product_name} ({len(results['images'])} 张图片)")
        return results

    def batch_generate(self, requests: list[ProductImageRequest]) -> list:
        """批量生成多个产品的图片集"""
        all_results = []
        for i, request in enumerate(requests):
            print(f"\n{'='*50}")
            print(f"Processing {i+1}/{len(requests)}: {request.product_name}")
            print(f"{'='*50}")

            try:
                results = self.generate_product_set(request)
                all_results.append(results)
            except Exception as e:
                self._log(f" 失败: {request.product_name} - {str(e)}")
                all_results.append({"product": request.product_name, "error": str(e)})

        # 生成批量报告
        self._generate_batch_report(all_results)
        return all_results

    def _generate_image(self, request, template, output_path, **kwargs):
        """生成单张图片（根据 method 选择不同的生成方式）"""
        prompt = generate_prompt(template, request.product_description, **kwargs)

        if self.method == "openai":
            return self._openai_generate(prompt, output_path)
        elif self.method == "comfyui":
            return self._comfyui_generate(prompt, request.source_image, output_path)
        else:
            raise ValueError(f"Unknown method: {self.method}")

    def _openai_generate(self, prompt, output_path):
        """GPT Image 2 生成"""
        response = client.images.generate(
            model="gpt-image-2",
            prompt=prompt["positive"],
            size="1024x1024",
            quality="hd",
            n=1
        )
        # 下载并保存
        import requests
        img_data = requests.get(response.data[0].url).content
        with open(output_path, "wb") as f:
            f.write(img_data)
        return output_path

    def _post_process_amazon(self, image_path):
        """Amazon 主图后处理"""
        output_path = image_path.replace(".jpg", "_amazon.jpg")
        create_amazon_main_image(image_path, output_path)
        return output_path

    def _log(self, message):
        timestamp = datetime.now().strftime("%H:%M:%S")
        self.log.append(f"[{timestamp}] {message}")
        print(f"[{timestamp}] {message}")

    def _generate_batch_report(self, results):
        """生成批量处理报告"""
        report = f"# 产品图批量生成报告\n\n"
        report += f"生成时间: {datetime.now().isoformat()}\n"
        report += f"总产品数: {len(results)}\n"
        report += f"成功: {sum(1 for r in results if 'error' not in r)}\n"
        report += f"失败: {sum(1 for r in results if 'error' in r)}\n\n"

        for r in results:
            if "error" in r:
                report += f" {r['product']}: {r['error']}\n"
            else:
                report += f" {r['product']}: {len(r['images'])} 张图片\n"

        with open(os.path.join(self.output_dir, "batch_report.md"), "w") as f:
            f.write(report)

# === 使用示例 ===
if __name__ == "__main__":
    pipeline = ProductImagePipeline(method="openai")

    products = [
        ProductImageRequest(
            product_name="Wireless Bluetooth Earbuds",
            product_description="premium wireless bluetooth earbuds with active noise cancellation, charging case, white color",
            brand_color="blue",
            target_platforms=["amazon", "shopify", "instagram"]
        ),
        ProductImageRequest(
            product_name="Stainless Steel Water Bottle",
            product_description="32oz stainless steel insulated water bottle, matte black, with bamboo lid",
            brand_color="green",
            target_platforms=["amazon", "shopify"]
        ),
        ProductImageRequest(
            product_name="Portable Neck Fan",
            product_description="portable bladeless neck fan with LED display, 3 speed settings, white and gray",
            brand_color="blue",
            target_platforms=["amazon", "instagram"]
        )
    ]

    results = pipeline.batch_generate(products)

5.2 A/B 测试图片变体

def generate_ab_test_variants(request: ProductImageRequest,
                              num_variants: int = 3) -> list:
    """为 A/B 测试生成多个主图变体"""
    variants = []

    # 变体 1：不同角度
    angles = ["front view centered", "45 degree angle", "slight top-down angle"]

    # 变体 2：不同光照
    lightings = ["soft studio lighting", "dramatic side lighting", "bright even lighting"]

    # 变体 3：不同构图
    compositions = [
        "product fills 85% of frame",
        "product fills 70% with more white space",
        "product with subtle shadow underneath"
    ]

    for i in range(num_variants):
        variant_prompt = (
            f"professional product photography, {request.product_description}, "
            f"{angles[i % len(angles)]}, {lightings[i % len(lightings)]}, "
            f"{compositions[i % len(compositions)]}, "
            f"pure white background, high resolution 8k"
        )

        img = generate_with_gpt_image(variant_prompt, f"variant_{i+1}.jpg")
        variants.append({
            "variant": i + 1,
            "angle": angles[i % len(angles)],
            "lighting": lightings[i % len(lightings)],
            "composition": compositions[i % len(compositions)],
            "image": img
        })

    return variants

6. AI 视频生成

6.1 产品视频类型

6.2 产品展示视频生成

# 概念代码：用 Runway API 生成产品展示视频
import runway

def generate_product_video(
    product_image: str,
    motion_prompt: str = "slow 360 degree rotation, studio lighting",
    duration: int = 4 # 秒
) -> str:
    """从产品图生成展示视频"""

    task = runway.image_to_video.create(
        model="gen3a_turbo",
        prompt_image=product_image,
        prompt_text=motion_prompt,
        duration=duration
    )

    # 等待生成完成
    task = runway.tasks.retrieve(task.id)
    while task.status != "SUCCEEDED":
        import time
        time.sleep(5)
        task = runway.tasks.retrieve(task.id)

    return task.output[0] # 视频 URL

7. 质量控制与合规

7.1 Amazon 图片合规检查

def check_amazon_compliance(image_path: str) -> dict:
    """检查图片是否符合 Amazon 要求"""
    img = Image.open(image_path)
    issues = []

    # 尺寸检查（最小 1000px）
    if min(img.size) < 1000:
        issues.append(f"尺寸不足: {img.size}，最小需要 1000x1000")

    # 白底检查（主图）
    pixels = list(img.getdata())
    corners = [pixels[0], pixels[img.width-1],
               pixels[-img.width], pixels[-1]]
    for i, corner in enumerate(corners):
        if not all(c > 240 for c in corner[:3]):
            issues.append(f"角落 {i} 不是纯白: {corner}")

    # 产品占比检查
    # ... (检查产品是否占画面 85%+)

    return {
        "compliant": len(issues) == 0,
        "issues": issues,
        "size": img.size,
        "format": img.format
    }

7.2 品牌一致性检查

8. 常见陷阱

8.1 用文生图做产品主图

文生图会重新「想象」你的产品，细节、比例、logo 都会偏。产品本身必须用图生图/图生视频，从真实产品照出发。这既是质量问题也是合规问题。

8.2 生成图不留元数据

哪张图是 AI 生成的、用什么工具、什么时候——这些在欧盟 AI 法案的透明度义务下是要能说清的。见 A6 §5。

8.3 批量生成不做人工筛

AI 出图的合格率没有想象中高，尤其是手、文字、反光、材质这几类。管道里必须有人工抽检环节，比例可以低但不能没有。

8.4 忽略各平台的图片规格

主图白底、尺寸下限、文字占比限制——各平台规则不同，生成时不带约束，最后要么被拒要么重做。

什么时候这套不管用

• 图要如实反映实物。 生成模型做不了“这个材质摸起来什么样“。主图和细节图涉及材质、颜色、尺寸的部分必须实拍，生成图只做场景和氛围。买家收到货和图不符，代价是退货加差评，某些平台还判定为误导性图片。
• 一套图要看起来是同一个产品。 生成有随机性，光影、色温、产品朝向每次都会飘。主图 + 场景 + 细节 + A+ 放在一起时，飘出来的不一致比单张图不好看更伤转化。要一致性就得固定种子、用同一参考图跑 img2img，或者实拍打底再做后期。
• 平台对 AI 内容的政策刚变过。 主图规则、AI 内容标注要求这几年一直在动，EU AI Act 的透明度条款直接管到这里（见 A6）。批量生成之前先确认目标平台当前的政策，别拿去年的理解跑今年的量。
• SKU 数量撑不起 Pipeline。 搭一套批量生成 + 后处理 + 合规检查的管道，成本要摊到几十上百个 SKU 上才划算。十几个 SKU 的话，直接用现成工具一张张生成加人工挑，比维护管道快。

9. 完成标志

• 搭建 ComfyUI 或选择 API 方案
• 为一个产品生成完整图片集（主图+场景图+信息图）
• 实现去背景+白底合成的自动化流程
• 构建批量生成 Pipeline（一次处理 5+ 产品）
• 通过 Amazon 图片合规检查
• 生成至少 1 个产品展示视频

(b8-ecommerce-dashboard.md) | Path 总览