How it works
The photo is scaled to 320 × 320 and passed through U²-Net, a salient-object segmentation network (the small 4.7 MB version), running in your browser with ONNX Runtime. It returns a soft mask of the main subject, which is scaled back up to your photo's full resolution with smooth interpolation, so edges stay natural instead of jagged. The edge slider moves the cut-off point of that mask: slide left to keep more hair and fur, right for a tighter cut that drops leftover background.
Example: a 3000 × 2000 product photo on a table becomes a 3000 × 2000 transparent PNG of just the product. Tick "crop to the subject" to trim the empty space around it, ready to drop into a thumbnail.
Getting clean cut-outs
It works best when the subject stands out from the background: a person, pet, product or logo with clear edges. Busy scenes with several subjects, or a subject the same color as the background, are harder. For emotes, cut out the face first, then resize it with the emote resizer. For thumbnails, put the cut-out on a bold color here, then check it in the thumbnail preview.
Frequently asked questions
Is my photo uploaded?
No. The AI model downloads to your browser once and runs there, so your photo never leaves your device.
Is the result full resolution?
Yes, up to 4096 px on the long side, with no watermark.
Why is part of my subject missing?
The model finds the most obvious subject. If it cut too much, slide the edge control left; if background is left over, slide it right.
Was this calculator helpful?