Many people try an AI cutout once, see a ring of the old background around the subject and conclude the tool is bad. Usually the photo simply did not give the model enough to work with. Cutout is a segmentation problem: the clearer the subject and the stronger the edge contrast, the cleaner the result.
1. At shooting time: separate subject and background
- Contrast: dark hair on a dark sofa, white shirt against a white wall -- the model cannot tell where the boundary is. Use a contrasting background, or move the subject further from it.
- Lighting: backlight blows out or blackens the edges and destroys edge information. Front or side light keeps the silhouette readable.
- Depth of field: a little background blur helps, but heavy blur mixes edges into the background and makes separation harder.
2. Choosing the photo: some images are inherently hard
Fine hair, sheer dresses, glassware, splashing water, products that match the background (white shoes on white) are all difficult samples. Accept that the result will need manual touch-up, or use a different shot with stronger contrast.
Resolution matters too: small images fail as well. Below 800 px on the long edge there are not enough edge pixels, so the cutout looks soft when enlarged. Start from 1500 px or more.
3. Settings: fine mode is not always better
"Standard" is fast and suits clean backgrounds. "Fine (hair)" runs alpha matting for more natural edges but takes longer and needs more GPU. Portraits, pets and plush products benefit from fine mode; hard-edged objects (boxes, phones, furniture) are better on standard, where fine mode can make straight edges fuzzy.
4. After the cutout: new background and compositing
A transparent PNG is only half the job; the destination decides the output:
- Dark web page: transparent PNG (the default from AI cutout);
- ID photo: fill with a solid colour via ID photo background;
- Product main image: white background via product white background;
- Print: convert to JPG on a white background so the printer does not render transparency as black.
5. Batch order matters
Filter first: drop obviously unusable images (too small, badly blurred, subject occluded), then batch. Otherwise one bad file slows the whole run and has to be redone. Test the settings on three images of different difficulty before running the full batch -- cheaper than rework.
6. Still a white halo on the edges
Halos usually come from a light background with semi-transparent edge pixels. Fix by eroding the mask by 1 px when filling the new background, or simply use a dark background. If it must be white and the halo is obvious, reshoot on a dark background.