FeyNoBg: Open-Source Background Removal Model Debuts
28 Jul 2026
A new open-source background removal model called FeyNoBg has been released alongside its training library, NoBg, on Hugging Face and GitHub. The release is being positioned as a state-of-the-art option for automatic background removal — a task widely used in e-commerce, photo editing, and content-creation tools.
What's new
FeyNoBg reportedly achieves the best published S-measure on 4 of 8 benchmarks tested, and comes within 2% of the leading result on the remaining four. The model architecture was expanded during development — parameter count grew from 222M to 263M, and the third stage of its feature extractor was expanded from 18 to 24 blocks.
Training relied on a curated dataset of 26.1K images pulled from 10 different sources, with each source capped at 4,000 images (anime images were capped lower, at 500) to manage dataset balance. The model was trained for 7,000 steps, with a final run adding 4,000 additional images from the S3OD dataset. Separately, the NoBg library was benchmarked against the original BiRefNet implementation at batch sizes of 1, 2, and 4, though the actual performance numbers from that comparison were not disclosed.
Both FeyNoBg and NoBg are released as open source, with NoBg specifically intended to encourage further community development in background removal.
Why founders should care
For early-stage founders building image-editing, e-commerce, or content-generation products, FeyNoBg's open-source availability likely lowers the barrier to adding background removal features without licensing a closed-source API. The reported benchmark performance suggests the model may be competitive with established background removal tools, though founders should treat these figures as a starting point rather than a guarantee — independent verification is probably warranted given the lack of disclosed benchmark names or metric methodology.
The parallel release of the NoBg training library is arguably the more interesting opportunity for technical teams: it opens the door to fine-tuning or extending the model for niche verticals (product photography, medical imaging, fashion, etc.) rather than relying solely on the pretrained weights.
That said, several gaps in the available information mean production planning should proceed cautiously. There's no published data on inference speed, latency, or hardware requirements, which could make capacity planning difficult until teams run their own tests. The license type for the open-source release also isn't specified, which may create ambiguity for commercial use — founders should confirm licensing terms before shipping FeyNoBg in a paid product.
What's missing
The report notes several open questions that could affect adoption decisions:
- The names of the 8 benchmarks and the exact definition of the "S-measure" metric aren't disclosed, making outside verification harder.
- No inference speed, latency, or hardware requirement data has been published.
- The actual performance numbers from the BiRefNet batch-size comparison aren't included.
- License terms for the open-source release aren't specified.
- Training compute cost, hardware, and time aren't detailed.
- There's no comparison to other popular commercial or open-source background removal tools beyond BiRefNet.
Bottom line
FeyNoBg looks like a promising, freely available addition to the background removal toolkit, with benchmark results that suggest it's competitive on at least half the tests reported. But founders evaluating it for production should plan to run their own performance and licensing checks before committing, given the current gaps in publicly available detail.