Background Remover

Cut the background out of any photo and download a transparent PNG. The AI runs entirely in your browser — your photo is never uploaded.

✂️ Click to choose a photo, or drag & drop it here
Works best on photos with a clear subject — people, products, pets. First use downloads the AI model (4.4 MB) once, then it is cached.

⬇ Download transparent PNG

How this is different

Most background-removal sites upload your photo to their servers. This tool ships the AI model (U²-Net, an open-source segmentation network) to your browser and does all the work on your device. That means:

Honest trade-off: on very fine details (wispy hair, fur against a busy background) heavyweight commercial services still win. For product photos, portraits, documents, and social-media cutouts, this does the job free and privately.

Why fine hair fails, and why a bigger photo does not fix it

This is the part every cutout tool leaves out, and it explains the one weakness in the table above better than "commercial services are better" does.

The network does not look at your photo at its own size. It works at a fixed 320 × 320 pixels — your image is shrunk to that, a mask is computed, and the mask is then scaled back up and used to cut the full-resolution original. Everything about the edge is decided at 320 px.

The consequence is worth stating plainly: uploading a larger photo does not produce a sharper cutout. A 4000 px wide image gets a mask with about twelve times less edge detail than its resolution implies, stretched to fit. That is the ceiling, and it is a property of the model rather than of this page. It is exactly why individual strands of hair, fur, and thin foliage come out soft or missing: at 320 px across a whole photograph, a hair is thinner than one pixel, so there is nothing for the network to keep.

Where this doesn't matter is the common case — a product on a table, a person against a wall, a pet, a document. Those have edges measured in tens of pixels even at 320, and the cutout is clean.

How to tell whether the cutout actually worked

Rather than leave you squinting at the preview, the tool reports what the mask itself says once it has run:

• Subject kept / background removed — the split, as a percentage of the frame. If almost everything was kept, or almost nothing was, the model failed to separate the two and the page says so.
• Uncertain — the share of the mask that came out partly transparent rather than clearly in or out. This is the useful number. Uncertainty collects exactly where cutouts go wrong, so an unusually high figure is a reliable early warning that the edges need checking before you use the file.

None of this requires trusting us: the percentages are counted from the mask that produced the image you are looking at.

What to use it for

Product listings (marketplaces prefer clean or transparent backgrounds), profile pictures, presentation graphics, stickers and memes, and prepping images for design tools — the transparent PNG drops straight into Canva, PowerPoint, or Photoshop.

Sources

• Qin et al. (2020), U²-Net: Going Deeper with Nested U-Structure for Salient Object Detection, Pattern Recognition 106 — the network this page runs. The variant shipped here is U²-Netₚ, the small version the authors published alongside the full model; it is the reason the download is 4.4 MB rather than 176 MB, and the reason the edge quality is what it is. The paper documents the 320 × 320 input size described above.
• MDN — HTMLCanvasElement.toBlob() — the browser API used to write the transparent PNG, which is why the alpha channel survives and why the file is produced on your device rather than on a server.

The 4.4 MB figure is measured from the file this site serves, not estimated. You can check it yourself in your browser's network tab on first load.

The guide that goes deeper

You might also need