Remove Image Background with AI
By Yasir Yaseen · Updated November 2026
Cut the subject out of a photo and save it as a transparent PNG. The model runs inside your browser tab, so the image is not sent to a server for processing.
Drop one image or a whole batch
or click to browse — PNG, JPG, WebP up to 20 MB each, up to 50 at once
Which photos cut out cleanly, and which do not
The tool uses a general-purpose segmentation model from the open-source @imgly/background-removal package. It decides, pixel by pixel, what counts as foreground. That works best when the subject is clearly different from what is behind it. The table below is guidance on what to expect, not a measured accuracy figure.
| Photo type | Expect | Watch for |
|---|---|---|
| Single person against a plain or out-of-focus wall | Usually clean edges | Check hair tips and the gap between arm and body |
| Product on a table or sweep of paper | Usually works well | Soft shadows under the object may be cut off or partly kept |
| Glass, clear plastic, jewellery with reflections | Unreliable | The model must decide whether see-through areas are subject or background; expect holes or a milky edge |
| Subject and background are the same colour (white mug on white) | Unreliable | Low contrast gives the model little to separate; reshoot against a contrasting surface if you can |
| Fine hair, fur, feathers against a busy scene | Rough edges likely | Fringes of the old background can survive; a neutral new background hides them better than a dark one |
| Crowds or several overlapping people | Unpredictable | The model picks what it treats as foreground; crop to one subject first |
When the cutout comes out rough
There are no sliders on this page, so the only levers are the input and what you do after. Match the symptom to the fix:
Edges look rough but the subject is intact
Retry from the original file, not a re-saved or screenshot copy. Heavy JPG compression adds blocky edges that the model reads as detail.
A chunk of the subject is missing
Crop tightly around the subject with the cropper and run it again, so the subject fills more of the frame.
A thin halo of the old background remains
Place the cutout on a background close to the old one in tone, or use Replace background with a soft, blurred scene.
Subject and background nearly match in colour
Reshoot or choose another frame. No setting on this page changes how the model judges contrast.
Why the output is always a PNG
Transparency needs an alpha channel. PNG and WebP have one; JPG does not. This tool writes PNG, named after your original file (portrait.jpg becomes portrait.png). A PNG of a photographic subject is usually larger than the source JPG, so run it through the compressor if you need to upload it somewhere with a size cap.
If a site rejects transparent files or shows them on black, that is the signal to fill the transparent area with a colour first. Add background does that, and Convert format can then produce a JPG.
First-use download, batches and phones
The first time you hover over or drag a file onto the drop area, the page starts fetching the model, which the code comments put at about 30 MB. On a slow mobile connection, that is the wait you see as "Loading AI model". After that it is cached by the browser and processing is local work on your CPU.
Selecting several files starts a batch of up to 50. Files are processed strictly one after another, not in parallel, because segmentation is memory-hungry and parallel runs can exhaust memory on phones. Each file shows its own status, and you can download results individually or as one ZIP. If a phone tab reloads partway through a large batch, work that was not yet downloaded is lost, so process big sets in smaller groups.
After a single image finishes, the Continue with bar can hand the PNG to Add background, Replace background, Resize, Compress or Make a PDF without uploading it again. The handoff happens on your device.
Questions about this tool
- Why does the first image take much longer than the rest?
- The cutout model is not part of the page. It is a download of roughly 30 MB that starts when you hover over or drag a file onto the drop area, and the first image has to wait for it. Once the browser has the files, later images skip that step. A private window or cleared site data means the download happens again.
- I selected 60 files and only 50 were processed. Where are the rest?
- A single run takes the first 50 files and ignores the remainder, so split larger sets into batches. Also note that if any file in the selection is not an image or is over 20 MB, the whole selection is rejected with an error and nothing is processed.
- Why is my downloaded file a PNG when I uploaded a JPG?
- A JPG cannot store transparency, so the result is always saved as a PNG with the same file name and a .png extension. If you need a JPG or WebP afterwards, for example with a white fill, use Add background first and then Convert format from the Continue with bar.
- What happens if one image in a batch fails?
- Images are processed one at a time, and each has its own status. A failure marks only that file as an error; the rest of the batch continues, and the ZIP contains the files that succeeded. The ZIP is stored without extra compression, since PNG files are already compressed.
- Is an animated GIF handled frame by frame?
- No. The file picker accepts GIF, but the browser decodes a still image from it, which in practice is the first frame. The output is a single PNG, not an animation.