Remove a photo background: when AI gets hair and glass wrong, and how to fix it

What the AI is deciding

A background remover answers one question for every pixel of your photo: how much of this pixel belongs to the subject? For most pixels the answer is all or nothing. Along the edge of hair, fur or a soft shadow the honest answer is “some”, and a good model says so with a partial value instead of a hard line. That partial edge is what lets a portrait sit on a new background without looking cut out with scissors.

Our Remove Background tool uses BiRefNet_lite, an open model published by its authors under the MIT licence. The page squeezes your whole photo to 512 by 512 pixels, the model draws its outline at that size, and the outline is scaled back up to your photo. Your own pixels are kept, up to 16 megapixels; only the mask comes from the small version. That one detail explains most of what goes wrong, and I come back to it below.

Which AI, and how I chose it

People ask which AI can remove a background for free. Several open models can, and before choosing one I measured them on 400 photos none of them had been trained on: 100 portraits from the PPM-100 set, the first 200 photos of AIM-500 (animals, people, plants, furniture, toys and transparent objects) and the first 100 of the ThinObject-5K test set.

The score is IoU, the overlap between the outline a model drew and the true outline, where 1.0 is a perfect match. I count a photo as a miss when its IoU falls under 0.8, which on screen means a visible chunk of the subject is gone or a visible chunk of background is left.

Model, measured on the same 400 photos	Overall IoU	Misses out of 400
BiRefNet_lite at 512, 8-bit (the one the site uses)	0.912	46
BiRefNet_lite at 1024	0.928	38
PP-MattingV2 (our passport tools’ people model)	0.822	103
U²-Net small (u2netp)	0.790	122
MODNet	0.759	139

The 1024 version scores a little higher, but its model file alone is 224 MB against 41 MB for ours, too much for a phone to download for one photo. I left out two well-known removers because of their licences: BRIA’s RMBG models are for non-commercial use only, and a site funded by ads is commercial; the in-browser library @imgly/background-removal is AGPL, a licence this site does not use.

Where it does well

On the photos most people bring, the misses almost disappear. Within AIM-500, the photos with one solid main subject (a person, a pet, a product, a chair) numbered 162, and the model missed one of them. By category it missed none of 72 animals, 36 portraits, 18 toys or 23 pieces of furniture. On the 100 PPM-100 portraits it scored 0.970, slightly ahead of the people-only model the passport tools use (0.965).

So for a product on a table, a dog on the grass or a person against a wall, expect a clean cut on the first try. The cases below are where a second look pays off.

Hair and fur

Hair is the case people worry about most, and it is handled better than you might expect: the model gives the outer strands partial values, so they blend into a new background. A halo is the more common complaint, a pale or coloured fringe around the head. That fringe is the old background’s colour still mixed into the edge pixels. The tool estimates the subject’s own colour in those pixels and takes the background’s colour out, so hair shot against a green wall does not glow green on white.

To check, use Hold to compare on the result, then look at the edge on a dark colour and on white, since a fringe that hides on one shows on the other. If a few strands were lost, Touch up with Restore and the small brush brings them back.

Glass, nets and veils

This is the honest weak spot. A wine glass, a plastic bottle or a bridal veil is partly see-through, so the right answer is “half subject, half background”, and the background you see through it changes with whatever was behind it. Among the 14 transparent objects in AIM-500, 11 were misses.

What works:

1. Shoot glass against a plain white or light background to start with. If the result will sit on white anyway, the parts seen through the glass already look right.
2. After the cut, open Touch up, choose Restore and paint over the missing parts of the glass with the small brush. Undo takes back a stroke that went too far.
3. For a transparent PNG of glass, expect to spend a minute on Touch up. The other models I measured did worse on see-through and fine detail: on AIM-500’s glass, net and fur photos, PP-MattingV2 scored 0.598 and MODNet 0.427, against 0.805 for this one.

Wires, strings and fine leaves

In the ThinObject-5K photos, 30 of 100 were misses. The 512-pixel view is the reason. A photo 4,000 pixels wide is squeezed almost eight times, so a cable or a twig thinner than about eight pixels in your photo is less than one pixel wide when the model looks at it. Plants in AIM-500 fared better (4 misses in 35), but the fine tips go first.

The fix follows from the cause: crop before you remove the background. A tighter crop with less empty space gives the subject more of the model’s 512 pixels, so thin parts survive. Crop Image does that on the device, and the cropped photo then goes into Remove Background. Touch up handles whatever is still missing.

Busy photos with no single subject

Remove Background looks for the main subject. A group of people, a crowded shelf or a street scene gives it nothing that clearly stands out, and in AIM-500’s photos of that kind it missed 9 of 14. When it finds no clear subject, the page says so and keeps the photo as it was. You can then paint the subject by hand with Touch up, or crop to the one thing you want and try again.

Logos, signatures and scans: skip the AI

A logo on paper or a signature on a white page is a different job. The AI is trained to find objects, and a signature is a few thin lines, the case it handles worst. Choose Logo, signature or drawing instead: it removes the paper’s exact colour, including inside letters such as o and e, with nothing to download. The Tolerance slider decides how pale a shade still counts as paper; raise it for grey scanner paper and lower it if pale ink starts to vanish. Save as PNG to keep the transparency.

A white background for a listing

Marketplaces want product photos on white, and Amazon is the strictest. In Amazon’s Seller Forums, Amazon’s own staff describe the main image background as pure white, RGB 255, 255, 255. Our White is exactly that (#FFFFFF), which is what White Background for Photos starts on. Off-white and Light grey are softer for a profile photo, but they are not pure white, so keep them away from a main listing image.

For a passport or visa photo, the background is only part of the rule. Several passport offices forbid editing the photo at all, so read the guide to passport photo sizes first, and use the Passport Photo Maker when the office allows a changed background.

Quality, file type and phones

The cut-out is saved at your photo’s own size up to 16 megapixels. A larger photo is made smaller to 16 megapixels first, since phones run short of memory beyond that. PNG and WebP keep the transparency; JPG cannot, so a transparent background saved as JPG comes out filled.

On an iPhone or an Android phone it runs in Safari or Chrome, with nothing to install. The first photo downloads the model and the runtime, about 55 MB, which stay in the browser’s storage for next time. Measured in the same runtime the page uses, on my M5 MacBook, the model takes about 3.6 to 3.8 seconds a photo once it is loaded, and the first photo of a visit takes longer while it warms up. A phone takes longer still; after a few photos the page tells you how long it usually takes on yours. The photo itself is never uploaded: the model, the edges and the saved file are all made on the device in your hand.