Feyn公司推出了FeyNoBg，一套用于自动背景移除的AI模型系统，并同步开源了支撑该模型的Python库NoBg [1]。FeyNoBg采用可解释性优先的训练方法，在来自10个数据集的26.1K多样化示例上进行训练 [1]。
该模型在八个基准测试中表现突出，其中四项取得最佳成绩，其余成绩均在最优值的2%以内 [1]。技术实现上，FeyNoBg通过扩展BiRefNet架构第三阶段的block数量（从18增至24）实现性能提升 [1]。NoBg库集成了图像matting模型的完整工作流，涵盖预处理、训练和评估环节，目前支持BiRefNet架构 [1]。
Feyn has unveiled FeyNoBg, an artificial intelligence model designed for automatic background removal, alongside the release of NoBg, an open-source Python library for training and deploying the system [1]. The model was developed using an interpretability-first training methodology and trained on a diverse dataset comprising 26.1 thousand examples drawn from ten different sources [1].
The FeyNoBg model demonstrated strong performance across multiple benchmarks, achieving top results on four of eight evaluation tests, with the remaining scores falling within 2 percent of the best performance [1]. The underlying technical architecture extends BiRefNet's third stage from 18 blocks to 24 blocks to enhance the model's capabilities [1]. The NoBg library integrates the complete workflow for image matting models, encompassing preprocessing, training, and evaluation stages, and currently supports the BiRefNet architecture [1].
The model is available through Hugging Face at https://huggingface.co/spaces/feyninc/feynobg, while the source code for the NoBg library can be accessed on GitHub at https://github.com/feyninc/nobg [1].