--- license: other license_name: mixed-mit-apache-2.0 license_link: https://huggingface.co/helloeditingmaster/master-mask-models/blob/main/LICENSE-SAM2-Apache-2.0.txt tags: - onnx - background-removal - image-segmentation - video-object-segmentation --- # Master Mask — model files ONNX model files downloaded by the **Master Mask** desktop app (macOS / Windows) on first launch. The app runs fully offline after this download. These files are not meant to be used on their own. `manifest.json` lists every file with its size and SHA-256; the app checks each download against it. ## Models and licenses | Folder | Model | License | Original source | |---|---|---|---| | `birefnet_lite`, `birefnet`, `birefnet_lite_2k`, `birefnet_lite_win`, `birefnet_win` | BiRefNet lite / full / lite-2K | MIT ([LICENSE-BiRefNet-MIT.txt](LICENSE-BiRefNet-MIT.txt)) | Peng Zheng et al., "Bilateral Reference for High-Resolution Dichotomous Image Segmentation" — https://github.com/ZhengPeng7/BiRefNet (weights: huggingface.co/ZhengPeng7/BiRefNet, BiRefNet_lite, BiRefNet_lite-2K) | | `sam2.1_tiny`, `sam2.1_small`, `sam2.1_base_plus`, `sam2.1_tiny_512`, `sam2.1_small_768` | SAM 2.1 hiera-tiny / small / base-plus | Apache-2.0 ([LICENSE-SAM2-Apache-2.0.txt](LICENSE-SAM2-Apache-2.0.txt)) | Meta AI, "SAM 2: Segment Anything in Images and Videos" — https://github.com/facebookresearch/sam2 | ## Changes from the originals - Converted from the official PyTorch weights to ONNX by Master Mask's export tool. - BiRefNet: two tensor reshapes are written differently (same maths) so the model runs on Apple's CoreML; `birefnet_lite_2k` is exported at two fixed input sizes (landscape and portrait). - BiRefNet for Windows (`birefnet_lite_win`, `birefnet_win`): the same models with each DeformConv (deformable convolution) written as standard operations (GridSample, MatMul), because the DirectML build of ONNX Runtime has no DeformConv. Same maths: largest output difference 1.7e-5. - SAM 2.1: split into image encoder, mask decoder, memory encoder and memory attention (for video tracking); `*_512` / `*_768` are the same weights exported at a lower working resolution. Small constant tensors are stored in `constants.npz` / `video_constants.npz`. The weights themselves are unchanged.