How BG Studio works

BG Studio removes and replaces image backgrounds without uploading your photo anywhere. The first time you use it, your browser downloads a small (~4.6 MB) neural network — U2-Net-p, an open salient-object detection model — and runs it locally with WebAssembly via onnxruntime-web. From then on, every pixel operation (matting, background compositing, brush refinement, color adjustment, export) happens on your own device.

The pipeline

Detect. Two on-device models share the work: MODNet, a portrait-matting network running at 512×512, handles people; U2-Net-p handles general objects at 320×320. Auto mode tries the portrait model first and falls back when its result is not confident (you can override this in the Edit section). A second pass then re-runs the chosen model on just the detected subject region, so the subject gets the model's full resolution.
Matte. The mask is upscaled to your photo's resolution, sharpened against the image's real edges with a guided filter, cleaned of stray specks, and applied as an alpha channel, producing the cutout.
Refine. Erase / restore brushes let you fix any spots the model missed, with undo and redo.
Compose. Choose a transparent, solid, gradient, image, or blurred background; move, scale, rotate, or mirror the subject; add a soft shadow or one-click color enhancement.
Export. Save as PNG (keeps transparency), JPEG, or WebP.

Privacy

There is no upload endpoint, no account, no analytics, and no server-side image processing. The only network requests are for this app's own static files. Very large photos are edited at up to 2400 px on the long side to stay fast in the browser. Closing the tab discards everything.

Limits worth knowing

The model is tuned for a clear main subject (people, products, animals). Busy scenes with several equally prominent objects may need the brushes.
Fine details like hair strands get a soft edge rather than a perfect matte — the Edge softness brush setting helps blend them.
Everything runs on your CPU; the first detection takes a few seconds while the model loads.

Credits & licences

MODNet portrait-matting weights — from the MODNet project (Ke et al., AAAI 2022), Apache License 2.0, fp16 ONNX export by Xenova, shipped unmodified.
U2-Net-p segmentation weights — from the U-2-Net project (Qin et al., 2020), Apache License 2.0, shipped unmodified. Provenance and checksums for both models: models/NOTICE.txt · licence text.
onnxruntime-web 1.27.0 — Microsoft, MIT licence, vendored.
ONNX export distributed by the rembg project (MIT).
Hosted on SkillSafe. BG Studio is an independent tool and is not affiliated with any other background-removal service.

Demo photo credits

The try-it demo photos are CC0 / public-domain images from Wikimedia Commons, resized for this app. No attribution is legally required for CC0/PD works; sources are credited anyway:

Couple — Happy Couple by freestocks.org (CC0).
Portrait — Brunette woman portrait by Christopher Campbell (CC0).
Hand — Setting sun in the West by Tim Graf (CC0).
Bird — Perched blue and orange bird by Vincent van Zalinge (CC0).
Rose — Vibrant red rose on gray by Aziz Acharki (CC0).

The built-in background gallery holds 38 CC0 photos from the same collection — three or four per category across Colorful, Beach, Sunset, Wall, Grass, Snow, Flowers, Street, Bokeh and Wood — each individually reviewed for backdrop suitability (no people, no readable text or brands, no dominant foreground subject) and resized for this app. Every photo and photographer is listed in models/NOTICE.txt.