Make an image background transparent

Make an image background transparent, then inspect the mask

The first result is an editable foreground mask, not a finished promise. U²-Net small looks for the visually salient subject in a 320 × 320 copy of the image. The soft grayscale result is then expanded to the decoded image dimensions, where white keeps pixels, black removes them and gray preserves partial transparency around edges.

Switch between Transparent result and Original before downloading. Hair against a similar-colored wall, bicycle spokes, gaps between arms, glass, smoke, shadows and several competing subjects are the places to inspect first. Paint Remove over background residue and Keep over subject detail that disappeared.

The automatic cutout is compact by design

The full U²-Net checkpoint is 176 MB. This page uses the authors’ 4.39 MB small checkpoint because a background tool that quietly downloads hundreds of megabytes is a poor web product, especially on a phone. The smaller model is fast enough to run locally and performed close to the full model on the authors’ published examples, but compact does not mean infallible.

The model is a 2020 salient-object detector, not a modern paid matting service trained on proprietary e-commerce data. That is why the page does not call the result “perfect AI,” does not export immediately after inference and keeps manual correction in the main workflow rather than hiding it in an advanced menu.

Keep and Remove change alpha, not the original colors

The correction brush edits only the mask. Remove paints transparency; Keep restores full opacity. Untouched automatic edges stay soft, so correcting one patch does not turn every strand of hair into a hard cut-paper outline. Undo removes the latest complete stroke, while Reset corrections returns to the model’s first mask.

The final operation draws the decoded source once, applies the reviewed mask as alpha and encodes a new PNG. PNG is required because JPEG has no alpha channel. The new file does not copy EXIF, GPS, embedded thumbnails or other source metadata through the browser canvas.

Inference and export stay on this device

The page shell does not preload machine-learning code. After Remove background locally, the browser fetches a versioned ONNX WebAssembly runtime and the U²-Net small model from CreatorValet’s own origin. A worker performs inference away from the page’s interaction thread. The selected image, normalized tensor, prediction, brush strokes and output PNG stay in this tab.

The runtime is MIT-licensed. The model is an ONNX export made directly from the upstream Apache-2.0 U2NETP checkpoint; its source revision, checkpoint hash, export settings and output hash travel beside the deployed assets. No remote IMG.LY, BRIA, remove.bg or other background-removal API is involved.

Why the model download is not compressed to one megabyte

An experimental uint8 export reduced the model to about 1.3 MB, but changed the masks materially. Dynamic variants were worse. The page ships the 4.39 MB float model because a smaller download is not a win if hair, holes and object boundaries become less reliable. Browser and HTTP compression may still reduce transfer overhead without changing model arithmetic.

Original dimensions are preserved, model detail is not invented

The PNG has the decoded source width and height, and manual strokes use full-image coordinates. The automatic edge began at 320 × 320, however. Expanding that mask cannot recover a strand, reflection or translucent boundary the model did not represent. Use the brushes for corrections that can be expressed as keep or remove; use a dedicated image editor when individual semi-transparent hairs need pixel-level matting.

Check the transparent PNG on two backgrounds

Open the download over white and over a dark color. A pale fringe hides on white, while a dark halo hides on black. Zoom into the outside contour, internal gaps and contact shadows. The receipt records dimensions, file size, alpha-weighted foreground share and manual stroke count; those facts prove what was exported, not that every semantic edge is correct.

If the next step is JPEG, first decide what should replace transparency. The PNG to JPG converter lets you choose a solid background instead of allowing transparent pixels to become an accidental black rectangle. The local-processing guide shows how to confirm that no image request leaves the browser.