Shreyash from Feyn introduces FeyNoBg, an automatic background removal model, and open-sources NoBg, the accompanying Python training library. FeyNoBg achieves state-of-the-art results across eight benchmarks, leading on four and nearing the leader on others. This performance is attributed to an interpretability-first training approach, expanding the third stage of the BiRefNet feature extractor from 18 to 24 blocks and training on 26.1K diverse examples.

The NoBg library addresses the common issue of isolated, incompatible codebases for image matting models. It offers a unified Python interface for training and running background removal models, starting with BiRefNet support and planning for more architectures. The library aims to streamline development and encourage further innovation in the field.

FeyNoBg utilizes an expanded architecture based on BiRefNet, growing from 222M to 263M parameters. This expansion, combined with training on a diverse dataset assembled from 10 sources, allows the model to excel at both identifying foreground subjects and precisely tracing their boundaries, even in complex scenarios like camouflage or fine structures like hair.

The FeyNoBg model is available on Hugging Face, and the NoBg library can be found on GitHub. The team is committed to answering questions and fostering development in automatic background removal.