Feyn Labs releases FeyNoBg and open-source NoBg training library

Feyn Labs released FeyNoBg, a background-removal model, and its open-source NoBg training library on July 21, claiming the model achieved the highest published S-measure on four of eight benchmarks and finished within 2% of the reported leader on the other four. Founder Shreyash Nigam said the release demonstrates the company's broader business of helping customers train custom models on proprietary data, run them under their own control, and improve them with production feedback.

Feyn Labs https://usefeyn.com/?ref=runtimewire founder Shreyash Nigam and FeyNoBg co-author Hafedh Hichri published FeyNoBg and its NoBg training library https://usefeyn.com/blog/feynobg/?ref=runtimewire on July 21, releasing the model weights and code behind Feyn Labs' latest experiment in specialist AI models. The release extends beyond automatic image cutouts. Nigam is using open-source models to demonstrate the broader business he and co-founder Bhavnick Singh Minhas are building: helping customers train models on proprietary data, run them under their own control, and improve them with production feedback. Feyn Labs' homepage frames the company around custom models that customers train on their own data, keep improving in production, and own as weights. Nigam arrived at that thesis after studying computer science at the University of Washington and working on cloud infrastructure, advertising personalization, and on-device AI at Google. Every reported https://www.every.io/blog-post/shreyash-nigam-and-bhavnick-minhas-are-building-chonkie-open-source-data-ingestion-for-any-ai-use-case?ref=runtimewire that context-window limits at Google pushed him to build Chonkie, an open-source document-chunking library, with Minhas, a childhood friend. The pair later took Feyn Labs through Y Combinator's Spring 2025 batch. Y Combinator currently lists https://www.ycombinator.com/companies/feyn?ref=runtimewire Feyn Labs as a three-person San Francisco operation. Adding capacity without discarding the base model FeyNoBg builds on BiRefNet, a high-resolution foreground-segmentation architecture with separate components for locating a subject and reconstructing its boundary. Feyn Labs expanded the third stage of BiRefNet's feature extractor from 18 blocks to 24, increasing the parameter count from 222 million to 263 million. The design choice reflects the problem Feyn Labs wanted to solve. A background-removal model must recognize the complete foreground object while preserving difficult boundaries around hair, fur, wires, blur, and other fine structures. The third feature-extraction stage retains enough spatial information to represent those edges while carrying enough context to identify the subject as a whole. Feyn Labs kept every compatible pretrained weight from BiRefNet and initialized only the six added blocks. That gave FeyNoBg extra capacity without forcing the full network to relearn the base model's existing representations. Training data was the second part of the bet. An initial run using only synthetic MaskFactory examples improved FeyNoBg on one benchmark while hurting it on another, according to Feyn Labs. Nigam and Hichri then assembled 26,110 images from 10 sources spanning crowded scenes, camouflage, portraits, high-resolution subjects, anime, and synthetic images. The largest sources contributed 4,000 examples each; Feyn Labs shuffled the combined set and trained for 7,000 steps. That mixture is central to Feyn Labs' argument. Background-removal systems can overfit toward either broad object recognition or precise boundaries when trained on narrow datasets. Feyn Labs says a varied training set produced a model that handled both more consistently. The benchmark claim has a defined boundary Feyn Labs says FeyNoBg achieved the highest published S-measure on four of eight benchmarks and finished within 2% of the reported leader on the other four. Its published results include scores of 0.981 on UHRSD-TE, compared with 0.957 for BiRefNet, and 0.983 on HRSOD-TE, compared with 0.961 for S3ODNet. Those figures are Feyn Labs' own evaluation and have not been independently validated in the material accompanying the release. The metric also deserves scrutiny. S-measure evaluates the structure and completeness of a predicted foreground mask. Feyn Labs converted segmentation masks and alpha mattes in its training data into binary foreground masks, meaning the matting datasets contributed detailed outlines without their soft-opacity labels. The results therefore make the strongest case for subject detection and structural boundary accuracy. They provide less evidence about partially transparent pixels, where traditional alpha-matting evaluations examine how accurately a model handles strands of hair, glass, smoke, and motion blur. Feyn Labs shows qualitative hair examples, but the field note does not publish a separate alpha-matting benchmark. That distinction does not make the benchmark irrelevant. Product photographs, portraits, vehicles, and many common cutout tasks depend heavily on complete foreground masks with clean shapes. It does mean the phrase "state of the art" should remain attached to the eight S-measure comparisons Feyn Labs selected rather than treated as a verdict across every background-removal workload. NoBg turns the model into a developer tool The more durable part of the release may be NoBg https://github.com/feyninc/nobg?ref=runtimewire , the Apache-2.0 Python library Feyn Labs used to train FeyNoBg. Developers can load models and processors through a shared interface, run batched inference, produce transparent PNGs at the source image's dimensions, fine-tune BiRefNet on image-mask pairs, and push resulting checkpoints to the Hugging Face Hub. Feyn Labs also published the FeyNoBg weights on Hugging Face https://huggingface.co/feyninc/FeyNobg?ref=runtimewire , allowing developers to run the model locally rather than send images to a hosted service. That separates the release from production APIs such as remove.bg https://www.remove.bg/api?ref=runtimewire , which packages background removal behind an HTTP request. NoBg gives developers the training machinery and the option to adapt a model to a specific image distribution. The open-source release fits a pattern. Feyn Labs' homepage also highlights Pulpie, a post about models for cleaning web pages, alongside Chonkie and other open-source work. Pulpie and FeyNoBg address different data types, but both act as public demonstrations of Feyn Labs' commercial thesis. Nigam wants to show that carefully selected training data, a task-specific architecture, and controlled evaluation can produce smaller specialist models that customers can own. NoBg makes that argument inspectable: the weights are downloadable, the training interface is public, and developers can test the claims on their own images. Feyn Labs has not disclosed revenue, customer count, or commercial pricing for its custom-model work. FeyNoBg also enters a category with mature hosted services and several open-weight alternatives. Feyn Labs' immediate advantage is the release itself. Nigam and Hichri have exposed enough of the training stack for other engineers to reproduce the workflow, challenge the benchmarks, and build specialized versions rather than accept another closed image-processing endpoint.