AI Background Remover — Erase Photo Backgrounds Online

In your browser. The U²-Net model is downloaded once on first use, cached locally, and the inference runs on every subsequent image inside the page itself.

About 5–15 seconds on a modern laptop while the ~44 MB model downloads and warms up. Every run after that takes 1–4 seconds per image because the model stays cached.

A PNG with a transparent alpha channel. Drop it onto a coloured background, swap in a new scene, or paste it straight into a slide.

Photos up to 25 MB. Very large images are auto-resized to a working resolution before inference and the alpha mask is upscaled back to the original dimensions on export.

For people, products and pets the result is usually clean. Hair, fur and motion-blurred edges are the hardest cases for any segmentation model — if the result needs cleanup, follow up with the manual brush in Manual Background Removal.

Chroma-key removal looks for one specific colour and treats every pixel near that colour as background — it works perfectly when you have a real green or blue screen behind your subject, and not at all otherwise. The AI Background Remover does not need any specific background colour. The U²-Net model has been trained to recognise foreground subjects regardless of what is behind them, so it works on photos shot in offices, bedrooms, gardens, and outdoors with no special setup.

The model runs the full image at its source resolution through a neural network that holds several intermediate tensors in memory simultaneously. A 50-megapixel photo needs roughly four times the working memory of a 12-megapixel one. 25 MB keeps the working set inside what mobile browsers can comfortably allocate without crashing the tab. For larger source files, run them through Resize Image first to bring the long edge down to ~3000 px — the cut-out quality is essentially identical at that resolution.

The current tool processes one photo at a time so the model fits comfortably in browser memory. To handle a folder, drop each file in turn — the tool stays open between runs and the model is already loaded, so each image after the first is fast.

Yes. PNG is the only common web format that supports a per-pixel alpha channel; JPEG cannot represent transparency, and WebP transparent is supported but has slightly worse browser-paste compatibility. If you need a smaller file, run the PNG through Compress Image afterwards — it will keep the alpha channel intact.

It handles flat-coloured illustrations and anime portraits reasonably well, especially when the subject is clearly outlined. Very thin line art on a white background sometimes confuses the segmenter into removing the lines along with the background. For pure line-art-on-white the older Make Semi-transparent or Transparent Background Maker tools may give a cleaner result via colour keying.

On first use, the browser downloads the ~44 MB ONNX model, the ONNX Runtime Web WebAssembly core, and a small JavaScript glue layer. Total initial cost is typically 5-15 seconds depending on your network. After that, all of those assets sit in your browser cache and the model also stays loaded in memory until you close the tab, so each subsequent image just pays the inference cost (1-4 seconds).

Always. The original file on your disk is never modified — the tool only ever reads bytes from it. The output PNG is a separate new file. If you want to try a different setting, just press Remove Background again with the same source.

It treats every person it can detect as foreground, so a group photo gets all subjects cut out together with whatever background is behind them removed. If you only want one person, crop the source first with Crop Image, then run the cropped version through this tool.