briaai / RMBG-2.0

## Overview RMBG-2.0 is an image segmentation model by BRIA AI for background removal. It predicts an 8-bit alpha matte, not a hard binary mask. Each pixel gets an opacity value from 0 to 255. That matte becomes a clean cutout with smooth edges. It helps you ship consistent product and portrait cutouts without manual masking. ## What you can build - eCommerce product cutouts with soft edges for hair, fur, and semi-transparent objects - Sticker assets for social posts, chat apps, and thumbnail designs - Batch background cleanup for catalog photos and marketplace listings - Preprocessing for design pipelines like background replacement, shadows, and compositing - Training data prep for vision tasks that need isolated foreground subjects ## Inputs - Upload 1 to 30 images for background removal in a single run. - Use high-resolution JPEG or PNG for best edge quality. - If you need transparency in the final asset, start from a PNG when possible. ## Outputs - Returns a background-removed image for each input. - The output includes an alpha channel derived from the predicted 8-bit matte. - Edges keep partial transparency, so cutouts blend better on new backgrounds. - Output size typically matches the original image dimensions. ## Recommended settings - Feed the highest-resolution source you have. Fine details improve with more pixels. - Prefer PNG inputs when you already have meaningful transparency. - If you need a hard cutout, post-process the alpha matte with a single threshold (0–255). ## Limitations - Very thin structures can fray. Expect issues with hair, fur, veils, and motion blur. - Semi-transparent materials can be tricky. Glass, smoke, and water may look imperfect. - Low-quality inputs reduce accuracy. Heavy JPEG artifacts and low contrast hurt edges. - Scanned photos, screenshots with compression, and busy patterns can create halos. - Model weights published for self-hosting are restricted by BRIA AI’s non-commercial license terms. ## Safety & compliance - BRIA AI trains this model on licensed data and markets it for safer commercial workflows. - Don’t use background removal to mislead viewers about real events or identities. - Respect image rights and privacy. Get consent for people in photos when required. - Check your license terms before any commercial deployment of self-hosted weights.