Remove an image background privately

Remove distracting backgrounds on your device while keeping the subject. FileGizmo runs an image-segmentation model locally, then lets you keep the transparent cutout or place the subject on a new background in the same panel. Refining and replacing work on one image; a queue of up to 20 processes sequentially into a single ZIP.

Quality mode uses BiRefNet Lite, an MIT-licensed segmentation model packaged with FileGizmo for local browser inference. The first run downloads 94 MB into browser storage, which takes longer than a conventional upload tool, though the photo itself is not sent with it. Fast mode is a 6 MB portrait-focused model for when download size matters more than edge quality, and it is the one chosen for you when the browser reports a metered or slow connection.

Batch mode processes images sequentially to control memory use. Manual refinement stays single-image because each mask needs review, and the background controls sit in that same panel, so a batch returns cutouts, not finished scenes.

Background removal predicts which pixels belong to the foreground; it does not understand the scene with human certainty. Fine flyaway hair, transparent glass, motion blur, smoke, fur, and foreground colors close to the background are the difficult cases, which is why real-photo QA runs loose hair and a translucent bottle rather than synthetic shapes.

Inspect the preview around hair, fingers, handles, and gaps inside the subject, because a cutout that looks clean on white can still show a halo on dark colors. Choose Refine edges when the automatic mask misses: Erase removes leftover background, Restore brings pixels back, and both edit the mask rather than the source pixels.

The brush runs from 4 to 80 pixels, and Hardness sets how abruptly its edge falls off. Precise paints under the cursor; Smart pulls a rough stroke toward a nearby strong edge. Feather and Smooth finish a mask that is accurate but cut too cleanly for the scene behind it. Zoom, Pan, hold Compare, and undo up to 20 strokes.

The Background fieldset decides what sits behind the cutout. Transparent leaves the alpha channel alone. A solid colour comes from presets, a picker, a hex field, or the eyedropper. Gradient blends two colours along an angle you set. Blur original rebuilds a backdrop from the photograph’s own pixels. Your image puts a photo from your device behind the subject, with Fill frame or Fit whole image and a five-way position control; it is never uploaded, and draft recovery does not restore it.

Despill, under Finish edges, pulls green contamination out of the subject’s own pixels, which is what a leafy or painted backdrop leaves in hair. Subject shadow casts the silhouette onto the new scene with offset, blur, and opacity; it only draws when something is behind the subject, so a transparent export never carries one.

The Download menu offers PNG or JPG at the source resolution. PNG is the answer whenever transparency has to survive, because only PNG stores an alpha channel, and it holds a new background equally well. JPG is the smaller file for a finished opaque picture, and a composed download carries a -new-background suffix, so it never overwrites the cutout.

Inference runs in a worker so the page stays responsive, but the model still uses real memory and CPU, and large photos go best on a desktop browser. The wait is a local-device cost, not an upload queue.

The model files are third-party software assets, not a hidden photo-processing API, and FileGizmo’s license inventory records their source and local deployment.

Use the result as a starting point for a listing, profile image, slide, or composite. A cutout is not evidence that an object was photographed on transparency, and a replaced background must not misrepresent a product, property, event, or location.