LibreWAFTl-flow-twins

WAFT (a2, Twins-SVT-Large encoder, 136.0 M parameters) optical flow weights, repackaged for LibreYOLO. Backbone variant twins of LibreWAFT size l.

Source

Derived from princeton-vl/WAFT at commit b152ff1, file waftv2-ckpts/twins/zero-shot.pth of the upstream Google Drive model zoo (zero-shot training schedule: TartanAir, then FlyingChairs, then FlyingThings3D). Copyright (c) 2025, Princeton Vision & Learning Lab. Licensed under the BSD 3-Clause License. The upstream model zoo publishes no separate weights license; LibreYOLO hosts the checkpoint under the repository's BSD 3-Clause License.

The encoder parameters come from Twins-SVT-Large (Meituan-AutoML/Twins via timm, Apache License 2.0).

Paper: Yihan Wang and Jia Deng, "WAFT: Warping-Alone Field Transforms for Optical Flow", ICLR 2026, arXiv:2506.21526.

Modifications

Checkpoint metadata wrapping only. Learned parameters and their names are unchanged. See weights/convert_waft_weights.py in the LibreYOLO source repository.

Usage

from libreyolo import LibreYOLO

model = LibreYOLO("LibreWAFTl-flow-twins.pt")
results = model(["frame_000.png", "frame_001.png"])  # N frames give N - 1 flows
flow = results[0].flow.data                           # (H, W, 2), pixels, (dx, dy)

License

BSD 3-Clause License (WAFT) together with the Apache License 2.0 (Twins-SVT-Large encoder). The combined LICENSE file contains both texts; see also NOTICE.