AI Background Removal Tutorial
Detailed guide, best practices, and FAQ
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.
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
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
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
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.