How U2-Net Becomes a Browser Cutout.

A mask is the model’s contribution

The automatic model does not deliver the entire editing experience. It predicts a mask from a prepared representation of the image. The application then scales that mask, combines it with the source, renders a preview and offers correction controls. Errors can arise in those surrounding steps as well as in the model prediction itself.

The U2-Net paper presents a nested U-shaped architecture for salient object detection. The portable variant used here is chosen for a manageable download and useful starting masks. We do not claim that it is the most accurate available model. Larger alternatives were screened by documented size, without a head-to-head quality trial.

The shipped graph is specific

The pinned weights are 4,574,861 bytes and accept a 320 × 320 model input. Those are properties of the actual graph used by the app, recorded in the source data. The application verifies the weight hash and attempts local WebGPU execution with a WASM fallback. A different U2-Net variant or converted graph could have different dimensions and behavior.

Keeping the graph pinned makes the public examples easier to interpret. If a model changes, the examples and evaluation need to be reconsidered. A model name alone is not enough to reproduce an output when weights, preprocessing and post-processing are unspecified.

Worked example: from camera photo to transparent file

The camera source is decoded locally. A smaller normalized representation goes to the model, which returns a foreground prediction. The app maps that prediction back to the source canvas and uses it as coverage. The original color pixels remain the visual subject; the mask determines which of them contribute to the output.

The editor then lets you restore a missing strap, erase a retained background area and choose the destination background. A transparent export preserves coverage, while an opaque export composites onto a chosen surface. The final file is therefore the result of model prediction plus application composition and any manual correction.

Full-size export has a precise meaning

The mask can be resized to the original dimensions, but interpolation does not create newly observed fine structure. A thin feature that disappeared in the model input may remain absent in the large export. This is why the site states the workspace cap and shows actual hair and glass limitations rather than promising perfect full-resolution matting.

Inspect difficult regions before downloading a batch. The automatic stage is a starting point for review. It is especially important to distinguish missing opaque details, which may be restored from the source, from transparent-material effects that require a different type of reconstruction.

Read the evaluation at its actual scope

Our examples prove that the shipped browser path produced real downloadable outputs for those licensed photographs. They do not establish a population-wide success rate or reproduce a paper benchmark. The qualitative captions identify what to inspect and where the result is limited.

Use a representative source from your own permitted workflow before committing to a repeated editing process. If the model consistently misses the kind of detail you need, a different capture setup or more specialized tool may be the better choice. Model size is only part of that decision.