An open-source, state-of-the-art AI model that automatically removes the background from an image and cuts the subject out onto transparency.
Last verified · 2026-07-27 · by Moe Ameen
FeyNoBg is an open-source model for automatic background removal, released by Feyn — a Y Combinator-backed developer-tools startup in San Francisco — and detailed in a technical write-up published on July 21, 2026 by Hafedh Hichri and Shreyash Nigam. The job it does is narrow and well-defined: given an image, it predicts an opacity value for every pixel, so the subject stays fully opaque while the background is made transparent. In practice that means you feed it a photo and get back a clean cutout of the person, product, or object with the backdrop knocked out — the alpha-matting step that sits at the front of most design, e-commerce, and thumbnail work.
Technically it is built on BiRefNet, the bilateral-reference architecture for high-resolution image segmentation. Feyn's contribution is to grow the network's third stage from 18 to 24 blocks — taking it from roughly 222M to 263M parameters (the model card lists it at about 0.3B) — while preserving the pre-trained weights during that expansion, so the extra capacity is added without throwing away what the base model already learned. The stated payoff is edge quality on the hard cases: fine structures like hair and fur, motion blur, and low-contrast, crowded, or camouflaged scenes where cheaper cutout tools leave halos or chew through detail. Feyn reports that across eight benchmarks the model posts the best published S-measure on four of them and lands within 2% of the leader on the rest, leading on high-resolution and dichotomous-segmentation sets.
It is released openly. The weights are on Hugging Face (under `nobg/FeyNobg`, in Safetensors), there is a hosted Hugging Face Space for testing an image in the browser, and Feyn also open-sourced NoBg — the Python library it used to train the model, which you can use to run FeyNoBg or train your own remover. You can load it through Hugging Face `transformers` (`AutoModelForImageSegmentation.from_pretrained("nobg/FeyNobg", trust_remote_code=True)`) or the NoBg library directly.
Two honest notes. First, FeyNoBg is fundamentally an image model — its own benchmarks include high-resolution image sets and a video-derived segmentation set (DAVIS-S), but "background removal for video" means running the model frame by frame, not a one-click video export. Second, it's a model and a library, not a consumer app: the demo Space takes a single image (PNG, JPEG, WebP, or HEIC up to about 5 MB), and doing anything at volume means writing code or wiring it into a pipeline. Treat specific numbers here as a snapshot from the launch write-up and check Feyn's pages for current details.
FeyNoBg is the sharpest possible version of one step: it hands you a clean subject on transparency. But a cutout is raw material, not a post — a transparent PNG of your product doesn't caption itself, sit inside a branded layout, become five carousel slides, or land on TikTok on Tuesday. That's the exact seam Kompozy closes. Run your shot through FeyNoBg, then bring the cutout into Kompozy as the subject of a Photo Post, a Carousel built brand-exact through HyperFrames, a Quote Graphic, or a face-locked Persona Photo — Kompozy drops it onto your brand background, layers on-style captions and price or offer text over it, and auto-reframes the frame to 9:16, 1:1, and 16:9 for each destination.
For product and e-commerce creators this is the whole workflow. One FeyNoBg cutout seeds a content unit: the product on a hero background for a Photo Post, the same cutout arranged across a Carousel of feature slides, a Quote Graphic pulling a review line over it, plus a Text Post and a short caption thread written in your voice through a Persona Brief. Then Autopilot and a per-post review pipeline schedule and publish the set across eight social platforms plus blog and email from one queue. FeyNoBg does the pixel-perfect matte; Kompozy turns that matte into a week of branded, captioned, scheduled posts.
FeyNoBg is an open-source AI model for automatic background removal, released by Feyn (a YC-backed developer-tools startup) in a July 21, 2026 write-up. It predicts a per-pixel opacity value so an image's subject stays opaque while the background becomes transparent, producing a clean cutout. It is built on the BiRefNet segmentation architecture, expanded to about 0.3B parameters, and is available on Hugging Face along with the open NoBg training library.
Yes — Feyn released the model openly. The weights are on Hugging Face (nobg/FeyNobg), there is a hosted Space to test an image in the browser, and the NoBg Python library used to train it is open source on GitHub, so you can run FeyNoBg or train your own remover. Check Feyn's pages for the exact license and any usage terms.
FeyNoBg is an image model. Its benchmarks include a video-derived segmentation set, but removing a background from video means running the model frame by frame through a pipeline — it is not a one-click video export. For a single photo, the Hugging Face Space accepts a PNG, JPEG, WebP, or HEIC up to about 5 MB.
FeyNoBg is the model, not a finished consumer service. It targets state-of-the-art edge quality on hard cases — hair, fur, motion blur, low-contrast and crowded scenes — and is meant to be self-hosted or wired into a pipeline via Hugging Face or the NoBg library. A hosted app like remove.bg is easier if you just want to drop in one image; FeyNoBg wins when you want control, high resolution, or to run cutouts at volume in your own stack.
FeyNoBg gives you the transparent cutout but does not publish it. Bring the cutout into Kompozy to build a Photo Post, brand-exact Carousel, Quote Graphic, or Persona Photo on top of it — HyperFrames composites it onto your brand background with captions — then reframe per platform and schedule and publish across eight social platforms plus blog and email from one queue.