AI Interactive Segmenter

What it does

Upload an image, click on the thing you want to keep, and the page returns that object on a transparent background. It uses MediaPipe's Interactive Segmenter — the MagicTouch model — which takes a small region of interest around your click and works out the full extent of the object it belongs to.

This is a different interaction from the automatic background remover. There, the model decides what the subject is. Here, you decide, one click at a time. That makes it useful for product shots, props, animals and anything else that is not a person.

The model and the interaction

MagicTouch is trained on the idea of a click: a tiny patch is enough to identify which object the user means, and the model then predicts a mask for the whole thing. The five-pixel region is only a seed. It is not a brush radius and it does not limit the result to a five-pixel area.

Only one region is passed per run. There is no way to add a second click to refine an edge, subtract a stray area or correct an over-eager selection. Each click replaces the previous result completely.

What appears on screen

After upload, the photo is drawn on the canvas at its natural dimensions, with a crosshair cursor to signal that it is clickable. The status line asks you to pick an object. Once you click, the model runs, the background is made transparent, and a preview of the cut-out appears in the side panel.

The canvas itself shows the cut-out at full size, which means the removed area takes on the page background rather than a checkerboard. The preview in the panel does the same at a smaller scale. To verify that the transparency is real, open the downloaded PNG in an image editor that displays an alpha checkerboard.

The download button appears only after a successful segmentation. The file keeps the original pixel dimensions and the original colours; only the alpha channel changes.

Step by step

1. Click the upload area and choose an image. Drag and drop is not supported.
2. Wait for the runtime and model to load; the panel names each stage. On later runs the model is already in memory.
3. Click directly on the object you want to keep. Aim for the centre of a solid area, not a thin edge.
4. Check the preview. If the wrong object was selected, click the right one and the result is replaced.
5. Download the PNG, or re-upload to start with a different image.

The Load AI Model button preloads the segmenter so the first click is fast. When it succeeds it reports "Model Ready" and disables itself.

Choosing a good click point

• Click inside the object, not on its outline. A point on a boundary is ambiguous between two regions.
• Avoid shadows and reflections. A dark shadow beside a product can be treated as part of it, or as a separate object.
• Pick a distinctive part. The centre of a large flat area is better than a small protrusion.
• Keep the object reasonably large. A tiny item in a wide shot gives the model little to work with.
• Watch for touching objects. Two overlapping items may be segmented as one region if the boundary is unclear.

Clicking the background

The model segments whatever region the click belongs to, not necessarily the subject. Click the background and you will get the background, with the object punched out — an inverse selection. That can be useful for removing a sky or a wall, but it surprises people who expect the tool to know which part is the subject.

There is no undo and no reset button. To recover, click the object again or re-upload the image.

Where it works well

• Product photography. A bottle, a shoe or a packaged item on a plain surface cuts out quickly.
• Props and objects. Furniture, tools, plants and food that the person-focused model would ignore.
• Animals. A clearly outlined pet usually segments well when it fills a decent part of the frame.
• Design mock-ups. Pull a single element out of a flat-lay for a layout.
• Quick masks. Produce a rough cut-out to test a composition before doing a careful job elsewhere.

Where it fails

Low contrast is the main problem. A white object on a white background gives the model almost no boundary to follow, and the mask bleeds outward or stops short. Transparent and reflective items — glass, mirrors, polished metal — are similarly difficult, because their edges are defined by what they reflect rather than by a stable colour.

Thin structures are fragile: chair legs, wires, plant stems and hair-like fibres may vanish or break into fragments. Very busy backgrounds with many distinct objects can also confuse the seed, especially if the click lands near a boundary.

Occlusion is handled by inference, which means a partly hidden object is completed with a guessed shape. The guess often looks plausible at a glance and wrong at full size.

Click accuracy depends on the canvas scale. The page maps your click into the image's own coordinates, and on a very large photo displayed much smaller, the sample can land a little away from the cursor. If a click keeps selecting the wrong thing, resize the image to roughly the displayed width before uploading.

What it does not do

No multi-point refinement, no brush, no eraser, no undo, no edge feathering, no background replacement, no batch processing, no video and no coordinate export. It does not remove an object and fill the hole; it keeps the object and removes everything else. It also does not label what it segmented — you get a mask, not a class name.

Privacy

The image is read locally, segmented in WebAssembly and drawn to a canvas. Nothing is uploaded, and no copy is retained after the page is closed. The runtime and model are fetched from public CDNs on first use, which shows the tool was opened but not what was in the picture.

Limits

One image per run, one click per result, one selected region at a time. No size limit is enforced by the page, but very large images consume more memory and take longer to decode. Output is always a PNG at the source resolution. A modern browser with WebAssembly and network access on first load is required.

Common mistakes

• Clicking an edge. Boundary points are ambiguous; click inside the object.
• Expecting to click several times to refine. Each click replaces the selection rather than adding to it.
• Clicking the background by accident. That selects the background and removes the subject.
• Saving as JPEG later. The transparency is lost when the PNG is flattened.
• Using a low-contrast photo. The model needs a visible boundary to follow.
• Judging the result in the small preview. Check the PNG at full size for edge defects.

Frequently Asked Questions