LibreFeyNobg supports weights released as FeyNobg:

Source: https://huggingface.co/feyninc/FeyNobg
        (reference library: https://github.com/feyninc/nobg)
License: Apache-2.0 (code and weights).
Copyright (c) 2026 Feyn Inc.

FeyNobg is built on BiRefNet:

Source: https://github.com/ZhengPeng7/BiRefNet
License: MIT.
Copyright (c) 2024 ZhengPeng (Peng Zheng).
Citation: Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen,
Wanli Ouyang, and Nicu Sebe. "Bilateral Reference for High-Resolution
Dichotomous Image Segmentation." CAAI Artificial Intelligence Research, 2024.

Architecture-sharing verdict (why this family has no nn code of its own):
FeyNobg keeps BiRefNet's module graph and only deepens the third Swin-L
stage from 18 to 24 blocks. This family therefore instantiates the birefnet
nn module with a family-local dimension table
(libreyolo/models/feynobg/nn.py) instead of duplicating the Swin tower and
bilateral-reference decoder. The upstream nobg library stores the same
architecture under Hugging-Face-style tensor names (split q/k/v attention,
indexed decoder ModuleLists); weights/convert_feynobg_weights.py remaps
those deterministically onto the original BiRefNet schema used in-tree,
with learned parameters unchanged. No code was copied from the nobg
library; the port is parity-gated against the nobg reference at fp32
(weights/parity_feynobg.py, max_abs_diff == 0).
