Cut Out a Subject and Keep the Edge Honest
Nobody has ever complained about the middle of a cutout. The subject comes out and the background goes, and that part has been solved for years. What people notice is a fringe of the old background still clinging to a shoulder, or hair that has turned into a solid shape with a jagged outline. Every meaningful decision in this job happens within about ten pixels of the boundary, which is why this page runs at four megapixels.
Use this without the search next time. Prathom Workbench puts Prathom's tools in your toolbar.
Add to Chrome — freeDrop an image here, or click to browse
Up to 25 MB. Uploaded temporarily, deleted after processing.
What it does
- Four megapixel working size for edge quality
- Soft alpha rather than a binary cutout
- Hair and fine detail kept as partial opacity
- Transparent PNG output
How to use Remove Background
- 1
Upload the largest version you have
Edge quality scales with resolution more than with anything else. A downscaled copy loses the pixels the boundary decision is made from.
- 2
Run the cutout
The image is resampled up to about four megapixels, a segmentation model produces a soft matte, and that matte becomes the alpha channel.
- 3
Check it over a color that clashes
Put the result on a background the opposite of the original. Fringing that is invisible on white is obvious on a strongly colored field.
How it works
Your image is first resampled up to about four megapixels — deliberately larger than most pages on this site use — and then passed to a segmentation model that produces a matte of the subject. That matte is inverted into an alpha channel and attached to the image, and the result is written as a PNG with real transparency.
There is no diffusion model and no sampler anywhere in the chain, so nothing in your picture is redrawn. The pixels in the output are the pixels you uploaded, resampled, with an opacity value attached to each one.
The problem lives at the boundary
Consider what a pixel at the edge of a subject actually contains.
A camera sensor integrates all the light that falls on one photosite. If half of that area was hair and half was the wall behind it, the recorded value is a mixture of the two. There is no arrangement of a single number that separates them again — the pixel is genuinely both.
For a rigid object this affects a thin line one or two pixels wide and nobody notices. For hair, fur, feathers, lace, smoke, or a glass edge it affects a band that can be dozens of pixels deep and contains most of what makes the subject look like itself.
The honest representation is therefore not a cutout at all. It is a matte with intermediate values, where a pixel that was sixty percent hair is sixty percent opaque. That is what this page produces, and it is why the result composites smoothly onto a new background instead of arriving with a cut edge.
Why the resolution is set high
Since the difficulty is that detail is finer than the pixel grid, the most effective intervention available is to make the grid finer.
Resampling does not add information about the scene. What it does is give the segmentation model more samples across the transition region, so the matte it produces has more room to express a gradual change from subject to background. The difference between a mask that steps from one to zero over two pixels and one that ramps over eight is very visible on hair and almost invisible on a bottle.
The cost is time and memory, which is why not every page here works this way. For a cutout it is worth it, because the edge is the deliverable.
Checking it properly
The universal mistake is checking a cutout against white.
Almost every fault this job produces — a fringe of the old background, a halo, an over-eroded edge that has eaten a millimeter of shoulder — is low contrast against white and therefore invisible. You approve it, place it on a dark brand color a week later, and the outline appears.
Open the PNG on something that clashes: if the original background was pale, view it on near-black; if it was a blue sky, view it on orange. Look along the whole silhouette, and pay particular attention to hair, to the gap between an arm and a body, and to anything transparent or reflective.
Thirty seconds there saves the version of this that gets noticed by somebody else.
Publication gate
This page ships once the workflow has been run against a rigid product on a plain surface, a portrait with flyaway hair against a bright window, a subject holding something with a gap through it, and a glass object, with every output composited over black, white, and a saturated color.
Examples
Product on a plain table
The easy case. A rigid object with a definite boundary is well determined at every pixel and comes out cleanly.
Portrait with loose hair
The hard case, and the reason for the resolution. Individual strands are thinner than a pixel, so those pixels are genuinely part background.
Frequently asked questions
Why does hair look wrong in most cutout tools?
Because a strand of hair is often narrower than a pixel. That pixel is a mixture of hair and whatever was behind it, so the correct answer is partial opacity rather than in or out. Tools that produce a hard yes-or-no mask have to pick one, and both choices are wrong at the same time.
What is the fringe of color around my subject?
The old background, still present in the edge pixels. Those pixels were always a blend, and keeping them fully opaque brings the blend along. It is invisible against a similar background and obvious against a different one, which is why the checking advice is to view it on a clashing color.
Does uploading a bigger file really help?
More than any other single factor. The boundary problem is that detail is finer than the pixel grid, so adding pixels directly adds information about the boundary. A strand that is half a pixel wide at one resolution is two pixels wide at four times the area, and two pixels can be described honestly.
Can I get a clean cutout from a screenshot?
Usually not a good one. Screenshots are small, often compressed, and compression damages exactly the high contrast edges the cutout depends on. You will get a result, and it will have the blocky boundary that the source already had. Find the original file if one exists.
Is the AI Remove Background tool free, and do I need an account?
It is free and there is nothing to register for. No account is created, no address is collected, and no reduced-quality preview is withheld for a plan that does not exist. The constraints that do apply come from the workflow itself rather than from billing: 25 MB in, a shared GPU queue, and a thirty-minute window to collect the output.
Further reading
- This PDF Kept Every Pixel and Still Came Out Looking WorstCompress a PDF and the obvious way to check what happened is to look at the resolution of the images inside. It is the number every tool reports and the one every comparison uses. It is also, on its own, close to useless — here is a file that kept every single pixel at full resolution and looked worse than one that threw three quarters of them away.
- YouTube Title Length: Where Titles Actually Get CutThe character limit is 100 and almost nobody hits it. The number that decides whether your title works is the point where each surface truncates, and those range from about 40 characters on a mobile home feed up to around 70 on desktop search. Which means the real skill is word order, not brevity.