LibreDPFlowm-flow-spring

Non-commercial weights. These DPFlow weights are licensed CC BY-NC-SA 4.0 by their author (trained within ptlflow). They are for non-commercial use only and are not covered by LibreYOLO's MIT license. You are responsible for complying with the weight license.

DPFlow (CVPR 2025) optical flow, repackaged for LibreYOLO's flow task. This checkpoint is the spring stage: fine-tuned on Spring from the Sintel stage (high-resolution benchmark).

DPFlow's dual-pyramid encoder picks its depth from the image diagonal (three levels around 1K, one more per doubling), so it runs at the native resolution of the frames and generalises to 4K and 8K inputs. LibreYOLO keeps that rule and the upstream test schedule (4 refinement iterations per level).

Usage

from libreyolo import LibreYOLO

model = LibreYOLO("LibreDPFlowm-flow-spring.pt")  # auto-downloads from this repo
r = model(["frame_000.png", "frame_001.png"])[0]
r.flow.data      # (H, W, 2) float32, pixels, (dx, dy) from frame 0 to frame 1
r.flow.to_rgb()  # Middlebury colour wheel

Source

Derived from hmorimitsu/ptlflow (ptlflow/models/dpflow, code at commit 43cdc2b5fc7228e7090009a66e8e28e0b48e31ba), weight file dpflow-spring-69bac7fa.ckpt from the weights1 release. Copyright 2025 Henrique Morimitsu.

Morimitsu, H., Zhu, X., Cesar Jr, R. M., Ji, X., and Yin, X.-C. "DPFlow: Adaptive Optical Flow Estimation with a Dual-Pyramid Framework." CVPR 2025.

Modifications

State-dict metadata wrapping only. Learned parameters are unchanged. See weights/convert_dpflow_weights.py in the LibreYOLO source repository.

Parity

Checked on an RTX 3090 against ptlflow's own DPFlow (commit 43cdc2b5) on 23 consecutive 640x360 frame pairs, fp32, TF32 off, deterministic cuDNN: max absolute difference 0.0 px, mean 0.0 px (bitwise identical).

License

Weights: CC BY-NC-SA 4.0. See the LICENSE and NOTICE files in this repository. The DPFlow code is Apache-2.0.