Batch Remove Background

Background removal here runs a segmentation model inside your browser tab through WebAssembly. The model decides, pixel by pixel, what is subject and what is not, and produces a transparent PNG. Nothing is uploaded, which for product photography under embargo or images of people is the difference between a usable tool and one you cannot put in a workflow.

The first run is slower than the rest, and that is expected rather than a fault. The model has to be downloaded and initialised before any image is processed; afterwards it stays in memory and subsequent images are much faster. A batch of fifty is therefore not fifty times the cost of one.

What it handles well is a clear subject against a background that differs from it: a person, a product on a surface, an object photographed for a catalogue. What it struggles with is predictable and worth knowing in advance. Fine hair and fur produce the hardest edges in the whole problem, because a single pixel genuinely contains both subject and background. Transparent and reflective objects, glassware in particular, confuse it because the background is visible through the subject. Low contrast between subject and background gives it little to work with, and a white product on a white surface is the classic failure.

Motion blur and shallow depth of field blur the boundary the model is trying to find, so a photograph with a soft edge produces a soft cut-out. This is not something a setting fixes; it is a property of the source image.

The output is a PNG with an alpha channel, and that has consequences downstream. Transparency is lost the moment the file is converted to JPEG, where it flattens to white or black. If the result is going into a system that does not accept PNG, the background needs to be replaced with a deliberate colour rather than discovered as an accident.

For work where the edge must be perfect, the honest expectation is that automatic removal gets most of the way and a person finishes the difficult parts. That is still a large saving on a batch, and it is a different claim from saying the result is final.