AIArticle
The Apache-2.0 training library matters more than the non-commercial weights it ships with.
Background removal is one of those unglamorous tasks that sits in more production pipelines than almost anything else in computer vision — e-commerce product shots, profile-picture flows, print-on-demand, video compositing. So when Feyn Labs shipped FeyNoBg last week claiming state-of-the-art results, it deserved a look. It got one on Hacker News, and the discussion landed exactly where it should have: not on the benchmarks, but on the license.
Here's the short version. FeyNoBg is a 263M-parameter matting model that leads on four of eight standard benchmarks and stays within 2% of the leader on the rest. It ships alongside NoBg, a genuinely useful Apache-2.0 training library. But the weights themselves are CC-BY-NC-4.0 — non-commercial — despite being a direct extension of the MIT-licensed BiRefNet. If that pattern sounds familiar, it's because Bria ran the identical play with RMBG-2.0: take BiRefNet, retrain it on better data, gate the result behind a non-commercial license. This is now the default business model of the background-removal niche, and it's worth understanding what it actually leaves on the table for you.
The model is mostly a data story #
Architecturally, FeyNoBg is a modest tweak: the team widened BiRefNet's third feature-extraction stage from 18 to 24 blocks, growing the model from 222M to 263M parameters while keeping the original pre-trained weights intact. That's it. The real work — and the real moat — is data curation. They assembled 26.1K images across ten public datasets (DIS5K, DUTS, COD10K, UHRSD, P3M-10K, HRSOD, CAMO, and others) with balanced sampling, and trained for just 7,000 steps.
The results are legitimately strong: S-measure of 0.981 on UHRSD-TE, 0.983 on HRSOD-TE, 0.961 on DIS5K, all category-leading. And to their credit, the writeup documents its own failure modes — adding the synthetic MaskFactory dataset improved camouflage segmentation (CAMO) while regressing DIS5K. That kind of honest trade-off reporting is rare in model releases, and HN commenters rightly singled it out as the best part of the post.
But notice what this implies. If a small team can beat the field with six extra blocks and a well-balanced 26K-image mix of publicly available datasets, the model itself isn't the scarce asset. The recipe is documented, the ingredients are downloadable, and the training run is short. Which brings us to the license.
The BiRefNet strip-mining problem #
BiRefNet — the 2024 dichotomous-segmentation architecture from Zheng et al. — is MIT-licensed, weights included. It has now spawned at least two prominent descendants that are not: Bria's RMBG-2.0 (source-available, commercial use requires a paid agreement) and FeyNoBg (CC-BY-NC-4.0). Nothing about this is improper — MIT explicitly permits relicensing derivatives, and whether model weights are copyrightable at all remains legally unsettled anyway. Feyn's founder was candid on HN about the reasoning: they've released MIT projects before and watched them get used without credit; non-commercial-with-attribution is their hedge.
Fair enough as a grievance. But as a distribution strategy, it's self-defeating, and one commenter nailed why: nobody wants to build on weights with an NC license, because it injects legal ambiguity into everything downstream. Is your internal tool "commercial"? Your freemium app? Your client work? Most engineering orgs answer these questions by picking a different model. That's precisely why rembg
— the MIT-licensed wrapper around U2-Net and friends — remains the default despite being visibly behind the frontier on hard cases like hair, fur, and translucency. The best weights in this niche keep pricing themselves out of the deployments that matter.
The library is the actual release #
The ironic part is that Feyn's durable contribution is the thing they gave away cleanly. Matting research has a packaging problem: every model lands as an isolated repo with its own preprocessing quirks, so comparing two models or fine-tuning one means writing adapters before you run a single experiment. NoBg fixes that with a transformers-style interface:
from nobg import AutoModel, AutoProcessor
model = AutoModel.from_pretrained("feyninc/FeyNobg").eval()
processor = AutoProcessor.from_pretrained("feyninc/FeyNobg")
It handles batched inference, half-precision, and — the part that matters — fine-tuning on your own image/mask pairs through the Hugging Face Trainer
API, with swappable losses (BiRefNet's composite loss, IoU, SSIM) and re-parameterization to change working resolution. One practical detail from the HN thread: for inputs larger than the model's native 1024×1024, they upscale the predicted opacity mask rather than tiling the image, which they found scales better than you'd expect.
What to actually do with this #
If you're building anything commercial, the move is not to license FeyNoBg — it's to use their Apache-2.0 tooling with unencumbered weights. pip install nobg
, load the original MIT BiRefNet checkpoint, and fine-tune on your own domain data (product photos, portraits, whatever your pipeline sees). FeyNoBg's own results are the proof this works: a targeted, balanced dataset in the tens of thousands of images and a short training run beat everything published. The datasets they used are public. The recipe is in their blog post. There's very little stopping you from reproducing 95% of the value in-house — on a single GPU, in an afternoon of training — with weights you fully own.
If you're non-commercial — research, personal tools, internal experiments you're confident stay internal — the FeyNobg weights on Hugging Face are the strongest published option right now, and the Space demo makes evaluation trivial. And if you just need "good enough" background removal with zero legal reading, rembg hasn't gone anywhere.
The bigger picture: background removal is commoditizing the same way speech-to-text did. The architectures are public, the training data is public, and the training runs are cheap. When the moat shrinks to "a well-curated 26K-image dataset," non-commercial licenses stop being a business model and start being a speed bump — one that a standardized training library, ironically the very one Feyn just open-sourced, makes trivially easy to route around. FeyNoBg the model will be leapfrogged within a couple of quarters. NoBg the library might be what everyone leapfrogs it with.
Sources & further reading #
FeyNoBg: A SOTA Model For Background Removal— usefeyn.com - Show HN: FeyNoBg - Automatic background removal model and training library— news.ycombinator.com - NoBg: background removal training and inference library— github.com - feyninc/FeyNobg model card— huggingface.co - briaai/RMBG-2.0 model card— huggingface.co
Mariana Souza· Senior Editor
Mariana covers the fast-moving world of machine learning and generative AI, with a particular focus on how these technologies are reshaping development workflows. When she isn't stress-testing the latest foundation models, she's usually at a local hackathon.
Discussion 0 #
No comments yet
Be the first to weigh in.