OmniShotCut v1.5 — MLX, 8-bit

uva-cv-lab/OmniShotCut_v1.5 converted for MLX by scripts/convert.py, for the Swift package OmniShotCutMLX:

import OmniShotCutMLX

let detector = try await OmniShotCut.pretrained("starkdmi/OmniShotCut-v1.5-mlx-8bit")
let segments = try await detector.segments(in: videoURL)

• Every linear but the query embedding is quantized to 8 bits in groups of 64; other matrices are fp16, biases and normalization parameters fp32. 67 MB; computed in fp32.
• Key names are the official ones, but backbone.0.body. is dropped, each attention's packed in_proj is split into q_proj/k_proj/v_proj, and convolutions are OHWI.

Against the official PyTorch code's 219 cuts on three Blender open movies, the package with these weights finds 214 on the same frame with the same labels: it misses one fade-in, relabels two cuts and places two a frame apart. The fp16 conversion, the package's default, finds 218. See the package's README for the test.

License

MIT, as is the original. Copyright (c) 2026 UVA Computer Vision Lab.

@article{wang2026omnishotcut,
  title={OmniShotCut: Holistic Relational Shot Boundary Detection with Shot-Query Transformer},
  author={Wang, Boyang and Xu, Guangyi and Zhang, Jiahui and Tang, Zhipeng and Cheng, Zezhou},
  journal={arXiv preprint arXiv:2604.24762},
  year={2026}
}