AI Background Remover v2
Remove backgrounds from any image — people, products, pets, objects — directly in your browser. BiRefNet lite is a general-purpose dichotomous image segmentation model, a substantial step up from portrait-only background tools.
Why an AI Tool That Runs In Your Browser
Most browser background removers are portrait-focused. They handle headshots well and struggle with everything else — a coffee mug picks up edges of the table, a cat keeps the chair behind it, a sneaker comes out with a halo where the floor meets the sole. The existing UDT MediaPipe-based tool is in that bucket: fast and reliable for selfies, frustrating for anything else. This tool replaces it for general use with BiRefNet lite, a bilateral reference network from Peng Zheng and collaborators that handles dichotomous image segmentation — the catch-all category covering general subject extraction, salient object detection, and camouflaged object detection. The lite variant trades a small amount of accuracy for a 339 MB to 85 MB (quantized) size that loads in a usable amount of time over a normal home connection. The full BiRefNet model is around a gigabyte and not currently practical for the browser. BiRefNet is released under the MIT license, the upstream model card is from ZhengPeng7 on Hugging Face, and the ONNX weights are mirrored by onnx-community. MIT permits commercial use, modification, and redistribution.
How AI Background Remover v2 Works
Click Load model on first visit. The browser downloads the quantized ONNX weights (about 85 MB) from the Hugging Face CDN and caches them in IndexedDB. Drop or pick an image from your device. The tool runs BiRefNet lite via the transformers.js background-removal pipeline, which produces a single-channel alpha matte at the same resolution as the input (up to a 1024 pixel internal working size). The matte is composited against the original image to produce a transparent-background PNG, which appears below the input with a download button. For complex edges — long hair, fur, leaves, semi-transparent fabric — BiRefNet is markedly better than the older MediaPipe-based tool. For high-volume portrait batch processing where speed matters more than edge quality, the original tool is still faster. Both stay live; this one is the new default for general work. The output is a 32-bit PNG with the alpha channel preserved, ready to drop into any compositing tool.
Frequently Asked Questions
What size is the BiRefNet lite model?+
BiRefNet lite is approximately 85 MB quantized, downloaded once from the Hugging Face CDN. The full-precision variant is closer to 339 MB; the quantized weights produce nearly the same masks at a fraction of the size. The browser caches the model in IndexedDB so later visits load in a few seconds.
Does my image get uploaded for background removal?+
No. After the model finishes downloading on first use, every background removal runs entirely in your browser. The image you pick stays on your device and is never sent to our servers or to a third-party API.
How does this compare to the existing UDT background remover?+
The existing tool is based on MediaPipe Selfie Segmentation, which is portrait-tuned. It is excellent on headshots and struggles on objects, pets, and products. BiRefNet lite is a general-purpose dichotomous image segmentation model and produces cleaner edges across all those categories at the cost of a larger model download and slower per-image inference.
What license does BiRefNet use?+
BiRefNet is released by Peng Zheng under the MIT license. MIT permits commercial use, modification, and redistribution. The ONNX weights are mirrored by the onnx-community organization on Hugging Face under the same license.
What kinds of subjects does it work well on?+
People, products, animals, plants, vehicles, furniture, food, and most everyday objects against a wider variety of backgrounds. It also handles complex edges like hair, fur, leaves, and translucent fabric noticeably better than portrait-only tools.
Are there input image size limits?+
The model resizes inputs to 1024 by 1024 internally before inference, then upscales the resulting matte to match the original image dimensions. Inputs of any reasonable size work; very large inputs (over about 4000 pixels on the longest side) use more memory and may run slowly on phones.
Why does the first removal take longer than later ones?+
The first run includes the model download (about 85 MB) and a warm-up pass through the network. Subsequent runs reuse the cached model and warmed-up runtime, so they only spend time on inference. On a modern laptop with WebGPU a warm removal takes 2 to 6 seconds; the WebAssembly fallback takes roughly 3x longer.
Can I batch-process multiple images?+
Not in the current UI — drag and drop processes one image at a time. The model itself supports batched inference, and adding a queue is on the v35 roadmap. For now, removing the background from a few dozen images means running them through one at a time.
Built by Derek Giordano · Part of Ultimate Design Tools
Privacy Policy · Terms of Service