PNG / JPG / WEBP
или кликните для выбора or click to browse
MAX 10MBTool guide
Your PNG, JPEG, or WebP is segmented by a neural network running inside the tab through onnxruntime-web: it labels object and background pixel by pixel and hands back a PNG with an alpha channel. The 6.5 or 44 MB model downloads once on first run and is served from browser cache afterwards, so every image after the first is processed with no network at all. Useful for marketplace product shots, avatars that need transparency, or a logo destined for a coloured backdrop. See also: trim the cut-out down to its edges, shrink the transparent PNG file size, trace a PNG logo into vector shapes.
RMBG-1.4 is 44 MB and trained on a wide range of subjects: gadgets, clothing, food, animals, objects with fiddly outlines. MODNet is only 6.5 MB and specialised in people — it is faster and more accurate on portraits and full-length shots, but it can leave background behind on an object with no person in frame. Start with MODNet for people, RMBG-1.4 for everything else.
On the first run the browser downloads the selected model's weights — 6.5 or 44 MB — and caches them. After that the network loads from cache in a fraction of a second and only the segmentation itself costs time. Switching models triggers one more download; the two models are cached independently.
No. The only network traffic is the model weights and the onnxruntime-web runtime. The image itself is read from the file, turned into a tensor, and stays in the tab's memory. Nothing carrying the picture leaves your machine, which makes the tool usable on work material you are not allowed to hand to a cloud service.
Try the other model first: MODNet usually catches loose strands on portraits, while RMBG-1.4 is cleaner on hard geometric edges. Turn on the compare checkbox to see exactly where the outline drifts. A larger source with a contrasting background helps too — a grey jumper against a grey wall is hard for any model because there is no real boundary there.
The export is a PNG with an alpha channel only, so it comes back genuinely transparent. Put that PNG over a white layer in any editor, or straight into a marketplace template, to get white. Keeping it transparent is more flexible: the same cutout then works over white, over a colour, or over another photo.
Yes. The file must be under 10 MB, which the tool enforces before anything starts, and only PNG, JPEG and WebP are accepted. Within that cap the ceiling is tab memory: images up to about 12-16 megapixels go through on an ordinary laptop, and on a phone stay under roughly 8. PNGs that already contain transparency are fine, the model simply segments what it can see.
Once a model has downloaded successfully, yes: the weights sit in browser cache and segmentation runs locally. Just do not clear the site data. After a cache wipe, or in a private window, the model has to be fetched again, and without a connection the tool will not start in that state.
Before: A sneaker on a wooden table, 2000x1500 JPEG with shadows at the edges
After: A 2000x1500 transparent PNG from RMBG-1.4 — the shoe drops straight onto a white template with no manual masking
Before: A staff photo shot against an office wall, 1200x1600 PNG
After: A PNG with an alpha channel from MODNet — shoulders and hairline preserved, the room removed entirely
Your rating and feedback help decide what to improve next.
