Back to Gap Map Infrastructure / Classic ML & computer vision

BiRefNet

open source / Overall score: 4.0(strong)

High-resolution dichotomous image segmentation model that separates a foreground object from its background, widely used for background removal and matting. Its Bilateral Reference design refines a coarse mask with high-resolution image patches and gradient cues. Checkpoints cover general use, a lighter model, 2K inputs, matting, portraits, camouflaged and salient objects, with training code in the repository.

Openness

5 medium confidence

5.0

weights: open(16 checkpoints on the Hub, none gated)
data: open(public segmentation benchmarks (DIS5K for the main model
code: open(training, loss and dataset code in the repository)
license: MIT(code and weights)

Code and weights are MIT-licensed, the training scripts are in the repository, and the main model was trained on DIS5K, a public dataset, with the variants adding other public matting and saliency sets. A few of those sets were not checked individually, which is why confidence is medium.

https://huggingface.co/api/models?author=ZhengPeng7&sort=downloads&limit=100&expand[]=downloads&expand[]=cardData&expand[]=gated recorded 2026-09-27
Sixteen ZhengPeng7 checkpoints, all gated false and tagged mit except BiRefNet-DIS5K-TR_TEs, which carries no license tag.
https://huggingface.co/ZhengPeng7/BiRefNet/raw/main/README.md recorded 2026-09-27
Model card: "license: mit"; "This BiRefNet for standard dichotomous image segmentation (DIS) is trained on **DIS-TR** and validated on **DIS-TEs and DIS-VD**."
https://raw.githubusercontent.com/ZhengPeng7/BiRefNet/main/LICENSE recorded 2026-09-27
LICENSE is the "MIT License", "Copyright (c) 2024 ZhengPeng".
https://raw.githubusercontent.com/ZhengPeng7/BiRefNet/main/README.md recorded 2026-09-27
README model zoo lists each checkpoint's training sets (DIS5K-TR; COD10K-TR and CAMO-TR; DUTS-TR, HRSOD-TR, UHRSD-TR; the general-use mix adds P3M-10k, AM-2k, HIM2K, Distinctions-646 and others), with the datasets "suggested to be downloaded from official pages".
https://ungh.cc/repos/ZhengPeng7/BiRefNet/files/main recorded 2026-09-27
Repository tree includes train.py, train.sh, dataset.py, loss.py, config.py and models/birefnet.py.
https://xuebinqin.github.io/dis/index.html recorded 2026-09-27
DIS project page: "Download (Newly Released Jul. 16, 2022) DIS5K Dataset (5,470)", split into DIS-TR (3,000), DIS-VD (470) and DIS-TE (2,000).

Adoption

4 high confidence

4.0

Hugging Face downloads over the trailing 30 days for the two declared checkpoints, the general model and the lite model; the other ZhengPeng7 checkpoints add comparatively little.

https://huggingface.co/api/models?author=ZhengPeng7&sort=downloads&limit=100&expand[]=downloads&expand[]=cardData&expand[]=gated recorded 2026-09-27
BiRefNet 798,071 and BiRefNet_lite 308,663 downloads in the trailing 30 days (1,106,734 together); all sixteen checkpoints sum to 1,273,908.

Capability

4 medium confidence

4.0

BiRefNet cuts the main subject out of any photo straight away, with no label set to retrain for, so a user's own images get a result from the shipped model directly. It does one task, foreground extraction, where Ultralytics spans many.

https://huggingface.co/ZhengPeng7/BiRefNet/raw/main/README.md recorded 2026-09-27
Model card gives an inference example that loads the checkpoint and predicts a foreground mask for an input image.
https://raw.githubusercontent.com/ZhengPeng7/BiRefNet/main/README.md recorded 2026-09-27
README describes "Bilateral Reference for High-Resolution Dichotomous Image Segmentation" and a model zoo of general-use, lite, 2K high-resolution, matting, portrait, COD and HRSOD checkpoints.

Verified 2026-09-27