RMBG is a background removal tool that runs entirely on the user's device. The service is free with an MIT license, requires no account, produces no watermarks, and imposes no artificial resolution cap. On first use the browser downloads approximately 98 megabytes of model weights from Hugging Face and an inference runtime from jsDelivr CDN; both are cached for subsequent visits.
How the BiRefNet-lite model works
The model loaded into the browser is BiRefNet-lite at 512px resolution, distributed under the MIT license. Provenance information is stored in models/manifest.json. Heavier non-commercial weights are supported through a bring-your-own approach and are never bundled with the application. The MIT license means the core model can be freely used, modified, and redistributed, which aligns with the tool's "no account, no credits" philosophy.
Privacy by design: on-device processing
Images never leave the user's device. The photograph is processed locally and its pixel data is never uploaded to any server. The first visit triggers two downloads: the inference runtime from jsDelivr and model weights from Hugging Face. After that, the browser cache typically reuses these files on return visits, though the browser may clear the cache depending on user settings or storage policies. WebGPU is the fastest backend; if unavailable the app falls back to CPU and indicates which backend is active. Startup time depends on connection speed and device capability.
Analytics and aggregate statistics
The site uses GoatCounter for page view tracking and fixed events related to model readiness, successful or failed cutouts, exports, and optional user votes of "Fine" or "Needs fixing." GoatCounter receives only the page path, referrer, browser type, approximate location, and screen-size metadata for aggregate statistics. The system does not receive image pixels, filenames, or image dimensions. GoatCounter describes its hosted service as cookieless and does not store persistent personal identifiers; its public dashboard is updated hourly and viewable by anyone.
User interface tools: erase, restore, guided, recompute
The editor provides several tools for refining cutouts. Brushes allow small slips to be painted and committed with Apply. Guided click enables one-tap selection of coherent regions. Drag-a-box recompute triggers when the AI misreads an entire area; the rest of the mask remains untouched. Every action is undoable, allowing iterative adjustments without losing prior work.
File formats: input and output
Supported input formats include PNG, JPG, WebP, AVIF, BMP, and GIF (first frame only); HEIC is not currently supported. Output options are PNG, WebP, JPEG, and AVIF where the browser can encode it. JPEG cannot store transparency, so the user must explicitly choose a background; the image is never flattened silently. Float alpha masks preserve hair and semi-transparency, and the original file is never mutated. Users can zoom to 100 percent to inspect edges and apply feathering or defringing to finish.
Batch processing workflow
Up to 24 images can be queued sequentially on the device. Each item in the queue supports per-item retry if processing fails. Any item can be opened in the full editor for individual refinement. Exported results can be bundled as a single ZIP archive containing all processed images.
Troubleshooting and error handling
If the model fails to start, retrying usually resolves the issue, as it is typically a download or GPU hiccup. Out-of-memory conditions suggest using a smaller image and closing other tabs. Undecodable files should be converted to a supported format before reattempting. When issues persist, the "Report an issue" link in the footer attaches diagnostics automatically; no image pixels are included in the report.
MIT license and model provenance
The core model BiRefNet-lite 512px is distributed under the MIT license, with provenance pinned in models/manifest.json. Heavier non-commercial weights are available on a bring-your-own basis only; they are never bundled automatically. This licensing model supports the tool's goal of being free, with no account required and no artificial restrictions on usage.