---
license: other
license_name: mixed
tags:
  - onnx
  - image-segmentation
  - depth-estimation
---

# Gamut — mask models

The on-device models behind Gamut's subject, sky and depth masks, kept here so every
copy of Gamut downloads exactly the same, checksum-verified files. Gamut downloads a
model only when someone chooses to use masks, and never sends a photograph anywhere:
all of this runs on the person's own computer.

These are unmodified copies of the files named below. Each keeps its own licence and
its authors' credit; nothing here is relicensed.

| File | What it does in Gamut | Model and authors | Copied from | Licence |
|---|---|---|---|---|
| `birefnet_lite.onnx` | Finds the subject | BiRefNet (lite), Peng Zheng et al. | `onnx-community/BiRefNet_lite-ONNX` (`onnx/model.onnx`), an ONNX conversion of `ZhengPeng7/BiRefNet_lite` | MIT — `LICENSE-BiRefNet.txt` |
| `sam2.1_hiera_tiny_encoder.onnx`, `sam2.1_hiera_tiny_decoder.onnx` | Picks out the subject someone points at | Segment Anything 2.1 (Hiera tiny), Meta | `vietanhdev/segment-anything-2.1-onnx-models` (`sam2.1_hiera_tiny_20260221.zip`), exported with samexporter by Viet-Anh Nguyen | Apache-2.0 — `LICENSE-Apache-2.0.txt` |
| `depth_anything_v2_small.onnx` | Near and far | Depth Anything V2 **Small**, Lihe Yang et al. | `onnx-community/depth-anything-v2-small` (`onnx/model.onnx`) | Apache-2.0 — `LICENSE-Apache-2.0.txt`. Only the Small model is Apache-2.0; the Base, Large and Giant models are CC-BY-NC-4.0 and are not used. |
| `skyseg.onnx` | Finds the sky | Sky segmentation by xiongzhu666 (U²-Net architecture, Qin et al.) | `JianyuanWang/skyseg`, from `xiongzhu666/Sky-Segmentation-and-Post-processing` | MIT — `LICENSE-skyseg.txt` |

## Papers

- Zheng et al., *Bilateral Reference for High-Resolution Dichotomous Image Segmentation*, CAAI AIR 2024.
- Ravi et al., *SAM 2: Segment Anything in Images and Videos*, ICLR 2025.
- Yang et al., *Depth Anything V2*, NeurIPS 2024.
- Qin et al., *U²-Net: Going Deeper with Nested U-Structure for Salient Object Detection*, Pattern Recognition 2020.