LibreMEMFOFm-flow

MEMFOF (Tartan-T-TSKH) three-frame optical flow, repackaged for LibreYOLO. This is upstream's MEMFOF-Tartan-T-TSKH, the authors' recommended checkpoint for real-world video, and LibreYOLO's default MEMFOF weights.

MEMFOF reads frames t-1, t and t+1 and predicts the backward and forward flow of frame t. In LibreYOLO it is the flow task family LibreMEMFOF: each result carries the forward flow from frame t to frame t+1 as Results.flow, an (H, W, 2) field in pixels on the original image canvas.

from libreyolo import LibreYOLO

model = LibreYOLO("LibreMEMFOFm-flow.pt")
results = model(["frame_000.png", "frame_001.png", "frame_002.png"])  # 3 frames in, 2 results out
flow = results[0].flow.data       # (H, W, 2), dx and dy in pixels
rgb = results[0].flow.to_rgb()    # Middlebury colour wheel

Source

Derived from msu-video-group/memfof (code at commit a51de9fc59c6fe20ba08e079372c7b583d58a712) and the checkpoint egorchistov/optical-flow-MEMFOF-Tartan-T-TSKH at revision 6c6c9aa3ad64f93aee8efbc2f7a6e4535814ee96. Copyright (c) 2024, Princeton Vision & Learning Lab. Copyright (c) 2025, MSU Graphics & Media Lab. Licensed under the BSD 3-Clause License.

Paper: Vladislav Bargatin, Egor Chistov, Alexander Yakovenko, Dmitriy Vatolin. "MEMFOF: High-Resolution Training for Memory-Efficient Multi-Frame Optical Flow Estimation." ICCV 2025. arXiv:2506.23151

Modifications

Metadata wrap only. The upstream model.safetensors tensors are unchanged (same keys, same values); they are stored in a LibreYOLO v1.0 checkpoint with task: flow, nc: 1, names: {0: flow}, model_family: memfof, size: m and variant: tskh. See weights/convert_memfof_weights.py in the LibreYOLO source repository. Inference defaults match upstream: fp32, 8 refinement iterations, native resolution with replicate padding to a multiple of 16.

License

BSD 3-Clause License. See the LICENSE and NOTICE files in this repository.

The weights' licence is not stated separately upstream: the upstream Hugging Face repository tags the checkpoint bsd-3-clause in its card metadata but ships no licence file or weight-licence text, so the licence is taken to be inherited from the code repository, whose BSD-3-Clause text is reproduced verbatim in LICENSE.

Citation

@article{bargatin2025memfof,
  title={MEMFOF: High-Resolution Training for Memory-Efficient Multi-Frame Optical Flow Estimation},
  author={Bargatin, Vladislav and Chistov, Egor and Yakovenko, Alexander and Vatolin, Dmitriy},
  journal={arXiv preprint arXiv:2506.23151},
  year={2025}
}