What changes
The result adds an alpha mask; fine hair, glass, shadows and similar foreground/background colours can be partially removed.
Remove an image background
Create a transparent PNG with a local portrait model. Your image never leaves the browser.
Runs locallyNo uploadYou control saved copies
Portrait model
Best for people and portraits. The first run loads about 20 MB of versioned model/runtime files from this website and caches them.
Fine hair, transparent objects and busy backgrounds may need manual cleanup in another editor. There is no mask-repair brush here. Device limits can downscale large outputs on desktop as well as phones.
Inference runs in a worker. Your image is never uploaded.
Remove a simple image background on your device
A self-hosted segmentation model estimates the foreground mask in a worker and writes transparency into a new PNG. No image is sent to a model API.
For a reliable result
- Choose a clear portrait with visible separation from its background; the model is portrait-focused.
- Start the local model and wait for the bounded worker pass.
- Inspect edges at full size before downloading the transparent PNG.
Check the download
Place the PNG over both a light and dark background to reveal halos or missing areas. Keep the source for manual correction if needed.
Know before processing
The first use loads a large public model from this site. Device limits can downscale output, including on desktop. There are no manual mask-repair controls; inspect the cutout before using it.
Quick answers
- Will it perfectly preserve hair, glass and shadows?
- Not always. Similar foreground and background colours can confuse the mask. Inspect the transparent PNG over both light and dark backgrounds before using it.
- Is my image sent to an AI service?
- No. The segmentation model is served with the website and runs in a disposable browser worker. Its first use can take longer while the local engine loads.
Make a portrait cutout and inspect what the mask removed
Background removal estimates which image areas belong to a person and makes the rest transparent. This portrait model suits a clearly separated person better than glass objects or crowded scenes. It creates a new PNG without changing the original.
Completion does not prove accuracy. Hair, clothing, shadows and held objects can be removed or retained incorrectly. Inspect the cutout against contrasting backgrounds before downloading a transparent or solid-colour version.
Choose the source and background for the intended use
| Your situation | Recommended approach |
|---|---|
| A person stands clearly apart from the background | Try the portrait model, then inspect the entire outline. Similar foreground and background colours can still confuse the mask. |
| You need a cutout for another design | Keep Transparent selected and download the PNG. Retain that transparent copy before flattening it onto a colour elsewhere. |
| You want a plain coloured background | After removal, choose White, Black, Soft gray or a Custom colour. These controls put colour behind the same cutout; they do not improve the model's edge decisions. |
| The image is for an application or identity document | Check the receiving authority's current rules before editing. A plain background does not establish that retouching, this crop or this image is acceptable. |
Compare a transparent result with a white-background copy
The illustrated-portrait.png sample is a synthetic 600 × 800 illustration, not a real portrait or an application-ready photo. It demonstrates the export and recolour workflow, not recognition quality on real hair or skin.
- Choose illustrated-portrait.png and select Remove background. Allow the versioned model and runtime files to load if this is the first run.
- Leave Transparent selected. Inspect the preview, confirm the Output dimensions and download the transparent PNG.
- Select White under Choose a background. Wait for the colour-applied message, then download the separate white-background PNG.
- Open both files and inspect the shirt: this synthetic example loses some foreground detail, so it is not a clean cutout and should not be used as an identity photo. Compare it over a contrasting background; select Transparent again to restore that version in the tool.
Expected result. The workflow produces a transparent PNG and a separate PNG composited over white at 600 × 800 pixels. Our sample loses some shirt detail as well as its background. That observed failure is why successful export must not be mistaken for an accurate foreground mask.
Check the downloaded result
- Inspect the whole subject, not only the face. Check hair tips, gaps between arms and body, accessories, clothing and anything the person is holding.
- Compare light and dark backgrounds. A pale fringe can disappear on white, while a dark retained border can disappear on black.
- Confirm the Background label before downloading. A solid-colour PNG is opaque; the PNG extension alone does not prove that a file contains transparency.
- Compare Original and Output dimensions. Device-dependent pixel caps can reduce a large image before export, including on desktop browsers.
Troubleshooting
- Hair or part of the subject disappears
The model has assigned a low-opacity mask to material that should remain. Busy backgrounds and fine or translucent details make separation difficult.
Try this: Try a clearer original with better separation, or use an editor with manual mask correction. This page has no erase or restore brush, and changing the background colour cannot recover a missing region.
- A halo or patch of the old background remains
The estimated edge contains background pixels or partly transparent mixtures of the subject and its surroundings.
Try this: Inspect at full size on contrasting colours. Keep the source for manual cleanup rather than assuming a colour that hides the halo has corrected it.
- The first run stalls or the model cannot load
The browser needs model and runtime assets from this website before local inference starts. The worker can also fail when required browser features or memory are unavailable.
Try this: Allow asset downloads, retry after a network failure, or try a smaller source in a supported browser. Local processing does not mean an uncached first run requires no network.
- A downloaded cutout looks like it has a white background
The viewer may display transparent areas on white, or White may have been selected before downloading.
Try this: Check the selected background and place the file over a coloured surface. Choose Transparent and download again when an actual alpha channel is needed.
Limits and compatibility
- One automatic mask
- The tool accepts one JPG, PNG, WebP or AVIF image and exports PNG. It does not offer batch processing, mask painting, subject selection or a guarantee for non-portrait objects.
- Local image handling
- The image is processed locally in a worker. Source and output pixel limits vary by device; save the original and reviewed output separately. Background removal is not secure redaction or proof of official photo compliance.
Your files and form values are processed on this device. Downloaded copies remain wherever you save them. See how to check local processing and our privacy policy for site data and optional services.
Common questions
- Does choosing another colour rerun the portrait model?
- No. Colour choices composite behind the existing transparent cutout. Transparent restores the cutout retained by the current session.
- Can I export the result as JPG?
- This page exports PNG. Use Image converter afterward if the destination requires JPG, remembering that JPG replaces transparency with an opaque background.
Continue your document workflow
Maintained by Shakti Prasad Ram. Found a mismatch? Report the tool, browser and steps without sending a sensitive document. Back to the tool