LibreFlowSeekt-flow
FlowSeek optical flow (ResNet-18 context and feature networks, Depth Anything V2 ViT-S (Small)
encoder, 4 refinement iterations), repackaged for LibreYOLO's flow task.
Source
Derived from mattpoggi/flowseek at commit
ba6407370bceeb96244cad4d60d37489f6eb471e, checkpoint flowseek_T_TartanCT_TSKH.pth (TartanAir, then Chairs and Things, then the
Sintel + KITTI + HD1K + Things mixed fine-tune: the generalist checkpoint upstream
evaluates on Spring and LayeredFlow).
Copyright (c) 2025 Matteo Poggi and Fabio Tosi (University of Bologna).
FlowSeek: Optical Flow Made Easier with Depth Foundation Models and Motion Bases, ICCV 2025.
The checkpoint embeds the frozen Depth Anything V2 ViT-S (Small) encoder (DepthAnything/Depth-Anything-V2), so this file is self-contained. FlowSeek builds on SEA-RAFT (BSD 3-Clause).
Upstream trains two networks. This size uses the same weights as
LibreFlowSeeks-flow
with a different iteration count, as in upstream config/eval/flowseek-T.json.
Modifications
Metadata wrap only (LibreYOLO checkpoint schema v1.0). Learned parameters are unchanged.
See weights/convert_flowseek_weights.py in the
LibreYOLO source repository.
Usage
from libreyolo import LibreYOLO
model = LibreYOLO("LibreFlowSeekt-flow.pt")
results = model(["frame_000.png", "frame_001.png"]) # N frames give N - 1 flows
flow = results[0].flow.data # (H, W, 2) float32, pixels, (dx, dy)
License
Apache License 2.0 (FlowSeek parameters, implied by the releasing repository; Depth Anything V2 Small encoder, declared upstream).