AI 抠图使用教程

详细使用指南、最佳实践与常见问题解答

使用场景

AI 抠图工具基于 ISNet 神经网络,完全在浏览器中运行推理。适用于电商商品图、人像证件照、社交媒体素材、动物/物体主体抠图等场景。无需上传图片,隐私安全。

Use Cases

AI Background Removal is powered by ISNet neural network running entirely in the browser. Ideal for e-commerce product photos, portrait IDs, social media assets, animal/object subjects. No uploads — privacy safe.

功能特点

ISNet 神经网络:业界领先的人像/商品抠图模型,效果接近商业级
双模型选择:快速(isnet_fp16,~40MB)/ 标准(isnet_quint8,~80MB)
输出格式:PNG(无损透明)/ WebP(体积更小,透明)
批量处理:支持多张同时处理,网格视图显示进度
ZIP 打包:批量结果一键打包下载
拖拽上传:支持拖拽图片到上传区
进度提示:模型下载与推理进度实时显示
本地推理:图片不上传服务器,完全在浏览器处理
IndexedDB 缓存:模型首次下载后缓存,二次使用秒载

Features

ISNet neural network: industry-leading model for portraits/products, near-commercial quality
Dual models: Fast (isnet_fp16, ~40MB) / Standard (isnet_quint8, ~80MB)
Output formats: PNG (lossless transparent) / WebP (smaller, transparent)
Batch processing: multiple images processed sequentially with grid progress view
ZIP packaging: batch results downloaded as a single ZIP
Drag-drop upload: drop images directly onto the upload area
Progress indicators: real-time model download and inference progress
Local inference: images never uploaded; processed in-browser
IndexedDB cache: model cached after first download, instant reuse

使用示例

示例 1：场景一:电商商品 — 上传商品照片,AI 自动移除背景,输出透明 PNG 用于商品主图

示例 2：场景二:人像证件照 — 上传人像照,移除杂乱背景,得到透明人像用于证件照合成

示例 3：场景三:批量处理 — 一次上传 10 张产品图,顺序处理后打包 ZIP 下载

示例 4：场景四:动物主体 — 上传宠物照片,AI 识别毛发边缘并精准抠图

Examples

Example 1: Scenario 1: E-commerce — upload product photo, AI removes background, export transparent PNG for main listing

Example 2: Scenario 2: Portrait ID — upload portrait, remove cluttered background, get transparent subject for ID composition

Example 3: Scenario 3: Batch — upload 10 product images, process sequentially, download all as ZIP

Example 4: Scenario 4: Animal subjects — upload pet photo, AI detects fur edges precisely

最佳实践

首次使用建议选"快速"模型预览效果,满意后再切"标准"模型输出最终结果
建议使用 Chrome 90+ / Edge 90+ / Firefox 88+(需支持 WebAssembly SIMD)
图片分辨率过大(>4000px)时建议先用"图片压缩"工具缩小后再抠图,可加速 50%
批量处理顺序执行以避免内存溢出,每张约 3-10 秒(取决于图片大小与硬件)
模型缓存后可断网使用,适合离线场景
复杂背景(头发、透明物体)建议用"标准"模型获得更好边缘

Best Practices

Use "Fast" model to preview first, then switch to "Standard" for final output
Requires Chrome 90+ / Edge 90+ / Firefox 88+ (WebAssembly SIMD support)
For images larger than 4000px, compress first to speed up by ~50%
Batch runs sequentially to avoid OOM; ~3-10s per image depending on size and hardware
Cached model works offline — great for disconnected scenarios
For complex backgrounds (hair, transparent objects), use "Standard" model for better edges

常见问题

为什么首次加载这么慢?

首次使用需要下载 AI 模型(快速 ~40MB / 标准 ~80MB)。模型从 CDN 加载并缓存到浏览器 IndexedDB,后续使用秒载,且可离线使用。

图片会上传到服务器吗?

不会。AI 模型与图片都在浏览器本地处理,F12 打开 Network 面板可验证无图片上传请求。完全符合隐私优先理念。

快速和标准模型有什么区别?

快速模型(isnet_fp16)使用半精度浮点,体积小、速度快,适合预览;标准模型(isnet_quint8)使用 8 位量化,精度更高,适合复杂边缘(如头发)的最终输出。

为什么批量处理是顺序执行而不是并发?

AI 推理需要大量内存(WebAssembly + ONNX Runtime),并发处理多张图片会导致浏览器内存溢出崩溃。顺序执行确保稳定性。

支持哪些浏览器?

需要支持 WebAssembly SIMD 的现代浏览器:Chrome 90+、Edge 90+、Firefox 88+、Safari 14.1+。IE 与旧版浏览器不支持。

FAQ

Why is the first load so slow?

First use downloads the AI model (Fast ~40MB / Standard ~80MB) from CDN and caches it in IndexedDB. Subsequent uses load instantly and work offline.

Are images uploaded to a server?

No. Both model and images are processed in-browser. Open DevTools → Network to verify zero image upload requests. Fully privacy-first.

What is the difference between Fast and Standard models?

Fast (isnet_fp16) uses half-precision floats — smaller and faster, good for previews. Standard (isnet_quint8) uses 8-bit quantization — higher accuracy, better for complex edges (e.g. hair) in final output.

Why does batch run sequentially instead of concurrently?

AI inference consumes significant memory (WebAssembly + ONNX Runtime). Concurrent processing would OOM the browser. Sequential execution ensures stability.

Which browsers are supported?

Requires a modern browser with WebAssembly SIMD: Chrome 90+, Edge 90+, Firefox 88+, Safari 14.1+. IE and older browsers are not supported.

← 立即使用工具返回首页

AI 抠图 使用教程