AI Human Remover From Image

What this page actually removes

The name says object remover, but the behaviour is narrower and more useful than that: it cuts a person out of a photo and makes everything the model considers background transparent. You upload an image, the page runs a segmentation model over it, and the result is a PNG with an alpha channel where the person remains and the surroundings are gone.

It does not erase a selected object, fill a hole with invented pixels, or let you brush over something you dislike. If you need a lamp removed from a room, this is the wrong tool. If you need a head-and-shoulders cut-out for a slide, a profile picture or a design mock-up, it does exactly that job.

The model and the pipeline

The Selfie Segmenter is a small two-class model: every pixel is scored as person or background. It was built for selfie-style framing — a face and upper body filling a reasonable part of the frame — and it is fast enough to run on a phone. The page uses the category mask directly, so there is no soft alpha gradient to work with. Each pixel is either kept or dropped.

What the result looks like

The canvas is resized to the photo's natural dimensions and the original is drawn onto it. The mask is then applied by zeroing the alpha channel of every background pixel, and the canvas is re-encoded as a PNG. Because PNG supports transparency, the cut-out survives the download intact; a JPEG would have flattened the empty area to white or black.

The preview in the side panel is the same image at a smaller size, shown against the page background. On a light theme the removed area looks white, on a dark theme it looks dark. That is transparency, not a fill colour. To confirm, open the downloaded file in an editor that shows a checkerboard behind transparent regions.

The download button appears only after a successful segmentation. If the model returns no mask, the panel reports a warning and no file is offered.

Step by step

1. Click the upload area and choose an image. Drag and drop is not supported.
2. On the first run, the browser fetches the MediaPipe bundle, the WebAssembly runtime and the segmenter weights. The panel names each stage.
3. Segmentation starts automatically once the image is decoded — there is no Process button.
4. Review the preview. If the cut-out is clean, press Download PNG.
5. To try another photo, upload again; the model stays loaded and the second run is much faster.

The Load AI Model button preloads the runtime if you want to avoid the initial wait. When it finishes it changes to a green "Model Ready" state and disables itself.

Where the edges hold up

• Clear separation. A person against a plain wall, a sky or a blurred background separates almost perfectly.
• Defined clothing. Jackets, shirts and uniforms with distinct outlines survive well.
• Front-facing portraits. The framing the model was trained on gives the most reliable mask.
• Consistent lighting. Even exposure on both subject and background avoids the model confusing a shadow for part of the person.

Where the edges fail

Hair is the classic weak point. Fine strands, flyaways and curls sit between the two classes, and a hard 0.5 cut-off either keeps a chunky halo or slices the outline short. The result is a visible jagged edge that looks like a low-quality manual selection.

Transparent or semi-transparent objects — glasses, veils, thin fabrics — cannot be represented by a binary mask at all. They will either disappear entirely or stay as opaque blocks. The same applies to motion blur around a moving arm or a swinging ponytail.

Low contrast between subject and background is another failure mode: a person in a beige coat against a beige wall may lose part of a shoulder. Busy backgrounds with people in them can also confuse the model, because it is looking for person-shaped regions everywhere, not just the subject you had in mind. Bystanders in frame will be kept along with the main subject.

Small subjects are unreliable. A full-body figure in a wide landscape occupies too few pixels for a clean outline, and the mask tends to swallow the legs or the feet.

Good uses

• Presentations and documents. Drop a speaker cut-out onto a coloured slide without a white box around them.
• Design mock-ups. Place a person on a poster or packaging comp.
• Profile images. Make a transparent avatar that sits well on any background.
• Thumbnails. Separate a presenter from a room and recompose the shot.
• Quick prototyping. Test a layout with a real cut-out before commissioning a proper mask.

What it cannot do

There is no background replacement, no colour fill, no blur-behind option and no crop tool. It will not remove a selected object, clone out a distraction, or generate a new background. There is no edge-refinement slider, no feathering, no decontamination of colour spill around the hair, and no manual brush to fix mistakes.

It also does not preserve soft transparency. Everything is binary at the pixel level, so a glass held in front of the face or a wisp of smoke will not survive. If you need those, a dedicated matting tool with trimap refinement is required.

Privacy

The photo is decoded and segmented in the browser. Nothing is uploaded, and no copy is stored after the tab is closed. The page does request the runtime and model weights from public CDNs on first use, which tells those servers the tool was opened, but the image itself never leaves the device. That makes it suitable for personal photos, provided you have the right to edit them.

Limits

One image at a time, one model, no batch processing, no webcam, no video, no undo and no adjustment controls. Output is PNG only, at the source resolution, which means a large photo produces a large file. Very large images take longer to decode and use more memory. The page needs a modern browser with WebAssembly and network access on first load.

Common mistakes

• Expecting object erasure. It removes the background, not a chosen item.
• Uploading a group photo. Every person-shaped region is kept, not just one.
• Judging edges in the small preview. Open the PNG at full size before deciding it is clean.
• Saving as JPEG afterwards. Flattening the PNG destroys the transparency.
• Using a low-contrast photo. Subject and background need to differ for a clean cut.
• Assuming hair will be perfect. Fine strands are where binary masks show their limits.

Frequently Asked Questions