How to fix a bad cutout

Three common fixes, tested on real failures from this site, with the numbers. Not all of them work.

Processing happens on our server. The upload and the result are deleted within 30 minutes, and you can delete them yourself the moment the PNG is ready.

The short answer: the fix that works best is a different photo: the same subject against a plain background that clearly differs from it. Cropping close to the subject sometimes helps with hair. Boosting contrast does not help and often makes things worse. Glass did not come out right in any test, so plan a manual edit for it.

The test

We took three photos that the model handled badly: a studio portrait where the hair came out thin, an old headshot where the model dropped most of the body, and a wine glass. Each one was uploaded again after one change, and the opacity in the problem area was measured. The changes were:

  • Tight crop: cut the photo down to the subject so it fills more of the frame.
  • Contrast boost: automatic levels, then contrast +40% and brightness +15%, the kind of edit a phone app does in one tap.
  • Different photo: a similar subject shot against a plain, contrasting background.

Uploading the same file again was not tested as a fix, because it cannot be one: the model gives the same pixels every time.

Hair: crop helped a bit, contrast ruined it

Hair area of the original studio portrait Photo
First cutout, hair thin and faint First try
Cutout after a tight crop, hair denser Tight crop
Cutout after a contrast boost, hair almost gone Contrast
Portrait hair, four versions Average opacity in the hair area: 49 out of 255 in the original result, 92 after a tight crop, 5 after the contrast boost. With boosted contrast the model kept the face and threw away almost all the hair: 82% of the area became fully transparent. Source photo: Ray Jones / Universal Pictures, 1944, Wikimedia Commons, public domain. Cutout: this site, unedited.
Lower part of an old headshot showing the shoulders and dress Photo
First cutout, shoulders mostly missing First try
Cutout after a contrast boost, shoulders gone Contrast
Headshot body, three versions The first result kept the head and made 92% of the body transparent. After the contrast boost it was 97%. Contrast made the same mistake worse instead of correcting it. Source photo: unknown photographer, via DPLA, Wikimedia Commons, public domain. Cutout: this site, unedited.

Why would cropping help? The model always works on a 1024 by 1024 copy of the photo. When the subject fills more of the frame, each strand of hair gets more working pixels, and the model can see more of it. Contrast does the opposite of what you would hope: it flattens the subtle tones between hair and background, which are exactly what the model uses to tell them apart.

Glass: nothing worked

Bowl of a wine glass against a grey background Photo
First cutout, bowl patchy First try
Cutout after a tight crop, grey background stuck inside the bowl Tight crop
Cutout after a contrast boost, nearly unchanged Contrast
Wine glass, four versions In the first result a quarter of the bowl was solid and 58% removed, with the rest half transparent. The tight crop made it worse in a new way: the grey background showing through the glass was baked into the bowl. The contrast boost changed only 3% of pixels. Source photo: freestocks.org, Wikimedia Commons, CC0. Cutout: this site, unedited.

A different photo: the fix that worked, mostly

Finally, two new photos of similar subjects, chosen because their backgrounds were plain and clearly different from the subject.

Woman with dark hair on a beach against a bright sky, original photo Photo
The woman cut out with her hair intact, over a checkerboard Result
Close-up of loose hair strands in the cutout, enlarged twice Hair, 2x
Dark hair against a bright sky A big improvement. Around the loose strands, 20% of pixels are solid and 23% partly transparent, so fine hair was kept as hair. Two small faults remain: a light haze on the left and a faint halo above the head. Source photo: Mario Klassen, Wikimedia Commons, CC0. Cutout: this site, unedited. Processing: 8.4 s.
Wine glass close up against a dark background, original photo Photo
The glass cut out with dark background inside the bowl, over a checkerboard Result
Close-up of the stem and foot, mostly missing Stem
Wine glass against a dark background Still wrong, differently. The bowl came out fully solid, including the dark background seen through it. The stem was 95% removed, and the foot is a blur. Source photo: Markus Spiske, Wikimedia Commons, CC0. Cutout: this site, unedited. Processing: 11.7 s.

The results in one table

FixHairBody droppedGlass
Upload againNo changeNo changeNo change
Tight cropBetter, not fixedNot testedWorse
Contrast boostMuch worseWorseNo change
Different photo, plain backgroundGoodNot testedStill wrong

This is a small test on a handful of photos, not a benchmark. It does match how the model works, though: it reacts to the photo you give it, not to edits that try to make the photo look easier.

What to do, in order

  1. STEP 01

    Look at the result over black and over white

    Many faults are only visible on one of them. Decide whether you actually have a problem for where the image will be used.

  2. STEP 02

    Reshoot or pick another photo if you can

    Plain background, clearly lighter or darker than the subject, soft light, nothing overlapping the edge.

  3. STEP 03

    Otherwise, try a tight crop

    Especially for hair and small subjects in a wide shot. Skip it for glass.

  4. STEP 04

    Finish by hand

    Open the PNG in an editor and paint back or erase the problem area. Small fixes take a minute; glass needs a manual outline.

Questions people actually ask

Will uploading the same photo again fix a bad result?
No. The model gives exactly the same output for the same file every time. Something about the input has to change: the crop, the background or the photo itself.
Does increasing contrast or sharpening help the model?
In our tests, no. Boosting brightness and contrast made two of three results clearly worse and did not change the third. The model was trained on ordinary photos, and an over-processed photo looks less like what it knows.
Is there any way to cut out a wine glass automatically?
Not reliably with this kind of model. Glass shows the background through it, so there is no clear line between object and background for the model to find. All four attempts on this page failed. For glass, a manual path in an editor, or a studio photo against black or white, is the realistic route.
What is the quickest manual fix for a small mistake?
Open the PNG in any editor with layer masks or an eraser, and paint back or remove the problem area. A missing bit of hair or a leftover patch of floor is usually a one minute job, much faster than trying to make the model get it right.

Written and maintained by Denis. Last updated: November 2, 2026