When to use it
If you have an image with a fairly tidy background and want to pull something out of it on its own, this tool is the least trouble: e-commerce images onto a white background, portraits for transparent avatars, logos pasted onto other artwork, pet photos as stickers. As long as the subject outline is clear and differs from the background in brightness or colour, one pass usually gives a usable result.
Conversely, there are several kinds of image it handles badly: fine hair, translucent glassware and thin gauze or mist, long-haired toys and animals, and subjects whose colour is close to the background (a white shirt against a white wall, a grey cat on a grey sofa). The reason is that the model judges from a 320×320 thumbnail and scales it back up; edges finer than that scale — a single hair, open lace — do not exist in the thumbnail at all and only become a blurred transition after enlargement.
When the cutout is not clean, do not keep clicking the button; running it again will not improve it. Do the two cheapest things first: crop the extra background so the subject fills more than half the frame, which makes the subject larger and the edges more accurate in the 320-pixel decision image; second, try a different background colour to view the result, since white or a dark background hides part of a leftover edge while a transparent background exposes every flaw.
How to choose the settings
- Subject share: 50%–70% is recommended. Below 30%, the subject occupies less than 100 pixels in the 320-pixel decision image and fine hair blurs away; cropping away the surrounding empty space is the single most effective step for better edges.
- Input resolution: a long edge of 800–1600 px is enough. Inference always runs at 320×320, so whether the original is 4000 px or 800 px the model sees the same detail, and the only difference is a larger output canvas — a 320 mask enlarged a dozen times makes edge steps more visible.
- Brightness difference between subject and background: aim for a difference of 80 or more on the 0–255 scale. Within 30 (light grey against white, dark blue against black) missed cuts and leftover edges increase noticeably, so fix the lighting on the original image first if you can.
- Output background: choose white first for hair and fluffy subjects, since semi-transparent edges are least noticeable on white; for dark subjects do not use black, where the edge is invisible; choose transparent only when you know the downstream background colour for certain.
- Export size: keep the long edge of a transparent PNG within 2000 px; 800–1000 px is enough for e-commerce main images and 512 px for round avatars. PNG is lossless, so doubling the dimensions roughly triples the file size without improving precision.
Compared with other approaches
Photoshop's Select Subject and the various cloud cutout services are clearly stronger on hard images: bigger models, plus hair-level transparency refinement, so glass, gauze and flyaway hair can come out usable. This tool uses the 4.6 MB lightweight model u2netp, and its accuracy ceiling is where it is. In exchange, the model is downloaded once and stays in the browser cache, so it works offline afterwards, and your image never passes through any server — unpublished product shots, ID documents and internal material can all be dropped in directly. The cost is that speed depends on your device: on an old computer or a phone, a single image takes a few seconds to a dozen seconds.
But for hard-edged images on a solid background (studio product shots on white, scanned logos), the traditional magic wand and chroma key tools are more controllable: you can specify the background colour, shrink the edge and work to the pixel. On such images the semantic model sometimes cuts away the shadow under the product, or bites out a chunk because a highlight looks too much like the white background.
Also: there is no edge shrinking, no threshold and no feathering here, and both the model and the resolution are fixed. It gives a roughly correct mask, and complex images almost always need a few brush strokes afterwards on an online editor.
Additional Q&A
The original is sharp, so why do the edges still look stepped?
The cutout does not look at the original resolution: the image is first scaled to 320×320 and fed to the model, the mask comes back at 320×320, and it is then scaled back to the original size. On a 4000 px image one mask pixel is spread over a dozen pixels, so diagonal edges become jagged.
How do I remove the ring of white edge around the subject on a transparent background?
Those edge pixels already carry a mix of the old background colour, and the transparency the model gives them is a middle value, so they show a white edge over a dark background. First try a white background, or one the same colour as the old background, to see the effect; removing it can only be done by hand along the edge.
Can a finished transparent PNG be printed directly or listed on an e-commerce platform?
In most cases no. Printing has no concept of transparency, and transparent areas come out as paper white or press black; many platforms require a white background for main images, and an uploaded transparent PNG gets filled by the system with a colour you did not choose. Switch to a white background before exporting.
Is there any hope for a same-colour background like a white shirt against a white wall?
There is a little room but no guarantee. First crop the extra background on the original and enlarge the subject and try again; if it still fails, reshoot against a background with stronger contrast. Hand-tracing the edge is the last resort and very time-consuming.