How to use the Background Remover
- Open a photo. Drop it onto the box, paste with Ctrl+V, or click to browse. The first time, the page downloads the AI model (about 4.7 MB) and the engine that runs it; a progress bar shows how far along it is.
- Let it find the subject. A few seconds later the background is gone and the subject sits on a checkerboard, which stands for transparency.
- Refine the edge. Mask threshold decides how confident the model must be before a pixel counts as subject — 0 keeps its soft mask, higher values tighten it. Edge feather softens the border.
- Fix mistakes by hand. Set Brush to Erase to paint away leftovers or Restore to bring back lost parts; change Brush size as needed. With a transparent background and the brush on, removed areas show faintly so you can see what to restore. Undo brush or Ctrl+Z steps back.
- Choose a background and save. Keep Transparent, fill with a Color, or use a Blurred photo of the original for a portrait-mode look. Download PNG saves the full-size result; Download mask saves the black-and-white matte.
How the cut-out is made
The model is U²-Net, published by Qin and colleagues in Pattern Recognition in 2020 for salient object detection — finding whatever in a picture the eye goes to first. The paper offers two sizes: the full network at 176.3 MB and a small version at 4.7 MB, 37.5 times lighter. This page uses the small one, because it downloads quickly and runs comfortably on a laptop or phone through ONNX Runtime Web, Microsoft’s open-source engine for running neural networks in the browser.
Everything happens on your device. The photo is shrunk to 320 × 320 pixels, its colors are normalized the way the network was trained, and the model returns a 320 × 320 map where bright means “subject”. That map is stretched back to the photo’s size with smooth interpolation and used as the alpha channel.
That last step explains most of what you’ll see. On a photo 4000 pixels wide, each pixel of the model’s map is stretched across about 12.5 photo pixels. Smooth outlines — a mug, a shoe, a person’s shoulder — come out clean, because interpolation is good at curves. A single strand of hair, a lace edge or the gap between bicycle spokes is smaller than one mask pixel, so the model can only guess at it. That is why the threshold, the feather and the brush are there.
The model doesn’t know what a “background” is in your sense; it learned which object stands out. A product on a plain table is an easy call. A group photo or a busy market stall is ambiguous, and it may keep the most prominent person or object rather than all of them.
What it handles well, and what needs the brush
| Subject | What to expect | What to adjust |
|---|---|---|
| Product on a plain surface | Clean edge, often ready to use | Threshold 50–60, feather 1 px |
| Portrait or pet, simple background | Good outline, soft hair | Threshold 30–40 for wispy hair, then feather 2–3 px |
| Painting photographed on a wall | Canvas edge usually found | Raise threshold to trim the shadow strip |
| Several people or objects | One may be dropped | Restore the missing one with a large brush |
| Glass, smoke, very thin objects | Partly or wholly missed | Mostly manual: Restore, then Erase |
| Subject the same color as the wall | Edges wander | Erase along the edge at a small brush size |
Download mask gives you a grayscale image where white is kept and black is removed. Drop it into Photoshop, Affinity Photo, GIMP or Krita as a layer mask and you can keep refining with full-resolution tools, without losing anything already done here.
Tips from practice
Shoot for the cut-out. A subject a step lighter or darker than its background, with a little space around it, beats any amount of brushwork. A sheet of colored card behind a small object works wonders.
Use threshold 0 for hair and fur. The soft mask keeps the semi-transparent fringe that makes hair look natural. Then use Erase only where background color shows through clearly.
Work in short strokes. Each stroke is one undo step, and the last 12 are kept. Fix one stretch of edge per stroke and a slip costs you seconds, not the whole cleanup.
Pick the background before judging the edge. A cut-out that looks perfect on the checkerboard can show a dark halo on white. Switch to Color and preview it on the color you’ll actually use — white for an Etsy listing, for example.
Finish the file for posting. Add a credit with the Watermark Adder, then shrink it with the Image Compressor for Portfolios — choose WebP or PNG there, because JPEG cannot store transparency.