AI Background Remover

AI Background Removal — U²-Net, edge refinement, and the hair problem

Background removal looks like a single operation but is actually three stacked problems: segmentation (which pixels belong to the subject), alpha blending (how opaque is each subject pixel), and edge refinement (how do we handle semi-transparent fringes like hair or glass). This page walks through how imagemochi's implementation solves each one.

The model — U²-Net Lite

U²-Net is a segmentation network originally proposed by Xuebin Qin and collaborators for salient object detection. The "U" in the name refers to a U-Net architecture — an encoder compresses spatial information into a bottleneck, a decoder reconstructs a full-resolution output, and skip connections between encoder and decoder layers preserve fine spatial detail. The "²" (squared) indicates a nested architecture where each encoder and decoder block is itself a U-shaped sub-network. The result is a model that's unusually good at edge precision relative to its parameter count.

We ship U²-Net Lite in two quantizations — 4.4 MB fp32 for maximum quality and a slightly smaller fp16 variant for devices that support half-precision WebGPU. The tool auto-selects based on your device capability. Both run via ONNX Runtime Web; on WebGPU-enabled browsers (Chrome, Edge, recent Safari) you get 5–10× speedup over the WASM fallback, but the WASM path still produces identical results, just slower.

Multi-scale inference and why it matters

A single-pass segmentation at the input resolution tends to miss small details — earrings, stray hair, background bokeh that shouldn't be foreground. Multi-scale inference runs the model at three input resolutions (full, half, quarter) and composites the masks together, letting small-scale features ride on the high-resolution pass and large-scale consistency come from the low-resolution pass. This is what makes it possible to remove backgrounds behind frizzy hair or glass without the mask collapsing into a silhouette.

Edge refinement — the hard part

Raw segmentation output is a binary mask: pixel is foreground (1) or background (0). Real objects have soft edges — hair strands, fabric fuzz, water droplets. The edge-refinement pass takes the binary mask and the original colour image, samples a narrow band around the mask boundary, and estimates a continuous alpha value (0.0–1.0) for each pixel in that band. We use a modified guided-filter approach: the colour image guides the alpha estimation, so the refined edge follows the actual subject silhouette rather than the approximate mask one.

The tool exposes two edge controls — "edge refine radius" (how wide the uncertainty band is, default 70 px) and "edge feather" (additional smoothing, default 0). Increase the radius for hair-heavy subjects or semi-transparent objects; keep it low for hard-edged subjects like products on a white background.

Use-case walkthroughs

Output options and colour management

The default output is PNG with a true alpha channel, ready for compositing. You can also export with a solid colour background (white for marketplace listings, chroma green for video keying) or a gradient fill. Colour management stays sRGB throughout — the model was trained on sRGB data and the output preserves the source's colour profile. If your input is in Adobe RGB or ProPhoto, we convert to sRGB on import and the output will be in sRGB. For wider-gamut workflows, export the alpha mask separately and composite in your colour-managed editor.

Common failure modes and fixes

• Background colour similar to subject. The model uses more than colour (it uses texture and structure), but a subject wearing a green shirt on a green backdrop will still be harder to segment. Fix: try the multi-scale mode, or manually paint out the problem region in a dedicated editor after export.
• Multiple subjects, only one wanted. U²-Net is a salient-object detector — it finds the "main" subject. For group photos or complex scenes, it will include everything salient. Fix: crop first to isolate your target, then run background removal.
• Motion blur. Motion-blurred edges confuse the segmentation. Fix: if you have a sharper frame from the same sequence, use it instead; otherwise, accept a slightly rough edge and refine manually.
• Subject touching the frame edge. The model sometimes includes background pixels near the frame because of convolution padding effects. Fix: add a few pixels of padding (canvas extension) to the input in Crop tool, then remove background, then re-crop.

Privacy and model provenance

The model weights are the official Xuebin Qin U²-Net release, quantized by us for browser deployment. Weights are downloaded once and cached in IndexedDB. Your images never leave your device — verify in DevTools Network tab if you like. Anonymous performance telemetry (time, device, mask area) is collected; image content is not.

Pro unlocks for background removal

Batch mode (drop a folder, get a ZIP of transparent PNGs back) is Pro-gated because it's compute-heavy. Free accounts get unlimited single-file processing, forever. The Pro 30-day history also remembers your preferred edge settings across sessions, which is handy if you shoot consistent subjects. See pricing.