Show HN: FeyNoBg – Automatic Background Removal Model and Training Library

FeyNoBg, a model demonstrating state-of-the-art (SOTA) performance in background removal, along with NoBg, the library used to train it, has been released.

FeyNoBg is based on the existing BiRefNet architecture, and to simultaneously enhance subject localization and boundary reconstruction capabilities, the model's third-stage blocks were expanded from 18 to 24, increasing the parameters from 222M to 263M.

Key features are as follows:

• High precision: It achieved the best S-measure in 4 out of 8 benchmarks, and shows performance differences within 2% on the rest.
• Boundary handling capability: It effectively combines image matting techniques to handle complex boundaries such as hair, fur, and thin lines.
• Open source: The model is available on Hugging Face, and the library for training and running it is available to anyone via GitHub.

The model is designed to enable precise subject separation even in high-resolution (4K, 8K) images.