About UNBG

Why UNBG exists

Removing a background has become a small, everyday task: a product photo for a listing, a profile picture, a logo on a slide. Most online tools do it by uploading your photo to a server. Many of those tools are good, but for a simple cut-out that means handing over a file that may be personal, confidential or not yet public.

Modern browsers can now run capable AI models directly on your computer or phone. UNBG uses that to do the whole job locally. Your image stays with you, and there is nothing to sign up for.

How we work

• One obvious job. UNBG is not a design suite. It removes backgrounds and gets out of the way.
• Private by design, not by promise. The architecture doesn't send images anywhere, and an automated test checks that before every release.
• Honest limits. We tell you when a file is too large, when your browser can't run the model, or when a result had to be reduced in size.
• Evidence over claims. Every public claim on this site is tracked with the test or source that supports it. We don't claim to be faster or better than anyone.
• Free during the beta. No product account, no credits, no watermark. This describes the current beta, not a promise about future pricing.

Who runs UNBG

UNBG is run by [operator name: to be confirmed].

Model and third-party licences

UNBG's application code is proprietary. It uses third-party software and model files with their own stated licences.

The AI model

Model

BiRefNet lite 1024 (fp16, WebGPU)

Files used

jiabins0303/birefnet-lite-1024-webgpu (onnx/model_fp16.onnx)

Version

dc4edd9f7623961aa5ae2b186c1b428f4ed38d6a

Based on

ZhengPeng7/BiRefNet_lite (MIT)

Licence declared by model publisher

MIT

Attribution

BiRefNet by Peng Zheng et al. (MIT, Copyright (c) 2024 ZhengPeng); ONNX conversion by jiabins0303.

Training data

The model's authors trained it partly on the DIS5K research dataset. Its terms of use restrict commercial use of the dataset. How those terms apply to a trained model and to its outputs has not been settled, and UNBG does not give legal advice about it.

Integrity

SHA-256 4059896039dfccb0f15b9080ff06d11d90e499449bb045e797055eb8901cf5f4, listed with every chunk in model-manifest.json

Model

BiRefNet lite 512 (fp16)

Files used

studioludens/birefnet-lite-512 (onnx/model_fp16.onnx)

Version

4a3c40c36c94093cc1e724d9ea428b8fa4b57dc7

Based on

ZhengPeng7/BiRefNet_lite (MIT)

Licence declared by model publisher

MIT

Attribution

BiRefNet by Peng Zheng et al. (MIT, Copyright (c) 2024 ZhengPeng); ONNX conversion by studioludens.

Training data

The model's authors trained it partly on the DIS5K research dataset. Its terms of use restrict commercial use of the dataset. How those terms apply to a trained model and to its outputs has not been settled, and UNBG does not give legal advice about it.

Integrity

SHA-256 eff9216bb2f9d3f023d9c2b7196845a7485739ab1f231593633e4d2344ffc516, listed with every chunk in model-manifest.json

Software

UNBG uses ONNX Runtime Web 1.30.0 to run the model in your browser, plus other open-source libraries. The complete list of third-party components and their licences is published in third-party-notices.txt.

Homepage example photograph

“Golden Retriever Carlos” by Dirk Vorderstraße, via Wikimedia Commons, licensed under CC BY 2.0. The transparent version was processed locally with UNBG. Both static images share the same dimensions and framing. This example does not run a model when you visit the homepage.

Contact

General questions and feedback: [contact email: to be confirmed]. Privacy questions: [privacy contact email: to be confirmed]. Security reports: [security contact email: to be confirmed] (see how to report a vulnerability).

If something went wrong in the app, include the diagnostic code shown on the error screen. It contains no image data. Please don't attach images you consider private: we never need them to help you.