BiRefNet Lite quality mode
Use the locally packaged model for stronger subject edges on people and products, with no runtime model download from a third-party CDN.
Remove backgrounds from up to 20 images locally, then put one cutout on a colour, gradient, blur, or your own photo and save PNG or JPG.
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 at a time; a batch of up to 20 comes back as pictures you can open one by one, or download together as a ZIP.
Quality mode uses BiRefNet Lite, an MIT-licensed segmentation model packaged with FileGizmo for local browser inference. The first run downloads up to 132 MB, the model and the runtime that runs it, into browser storage: a device with WebGPU gets a 1024-pixel version that runs on the graphics card, and any other device a 512-pixel version on the processor, which takes longer than a conventional upload tool, though the photo itself is not sent with it. People only is a 29 MB first download with a smaller portrait model, for photos of people when download size matters more than edge quality; it does not cut out products, logos or signatures. Nothing downloads until you give it an image, so choose the model first. The progress line shows how much of a first download has arrived, and Cancel stops a run at any point.
Batch mode processes images one after another to control memory use, and Add more puts further images on the end of a finished batch without redoing the rest. Each result opens in the editor on its own, and what you download from there replaces it in the ZIP.
Background removal predicts which pixels belong to the foreground. Fine flyaway hair, glass, motion blur, smoke, fur, and colors close to the background are the difficult cases; the benchmark shows where each mode stands.
Check hair, fingers, handles, and gaps inside the subject, because a cutout clean on white can show a halo on dark colors. Open the Edges tab 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, compare with the Original slider, and undo up to 20 strokes, or use E, R, the bracket keys and Ctrl+Z.
The Background tab 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 Size tab saves PNG, WebP, JPG or the mask alone, at the source resolution unless you pick a size. PNG and WebP keep transparency; JPG is the smaller file for a finished opaque picture, and a composed download carries a -new-background suffix.
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.
Drop, paste, or choose up to 20 JPG, PNG, WebP, or AVIF images
Watch the local model cut each one out, starting as soon as they arrive
Download the image or the ZIP, or open one to add a background and refine its edges
Use the locally packaged model for stronger subject edges on people and products, with no runtime model download from a third-party CDN.
Correct the automatic mask with soft-edged strokes, tune hardness, feather, and smoothing, then judge the edges against a grid, white, or black preview.
Place the subject on a solid colour, a two-colour gradient, a blurred copy of the original, or a photo from your device, then add despill and a shadow and export PNG or JPG.
Processing runs on this device. Open the browser’s Network panel while the tool works and you will see no file upload request to FileGizmo. Temporary previews use local browser URLs, and closing or resetting the page releases them instead of leaving a server copy behind.
Use the tool without a daily task counter, account wall, output watermark, or artificial upload cap. FileGizmo does not meter a transfer it never receives. Available memory, processor speed, browser canvas limits, and the file format itself set the practical limit.
Use a current browser on desktop, Android, iPhone, or iPad without installing an app. Once this tool and its required assets load, the cached workflow works without a connection. Large jobs are usually more comfortable on a desktop with additional memory.
No. The model and processing run in your browser; neither your photo nor a background you supply is uploaded.
The browser must initialize the local machine-learning model. Later runs reuse cached assets.
People and clearly separated products with distinct edges generally produce the cleanest masks.
One image downloads as PNG or JPG, set on the Background tab. PNG keeps transparency and any new background; JPG is opaque and whitens a still-transparent one. A batch downloads as a ZIP of transparent PNGs.
Up to 20 JPG, PNG, WebP, or AVIF images per free batch, and larger sets can be split across batches. A batch returns cutouts, and any one of them can be opened to give it a background.