FreeFlow tops Sintel, KITTI, Spring without flow-specific parts

FreeFlow, a hierarchical transformer with no correlation volumes, no feature warping and no iterative refinement, posts state-of-the-art optical-flow scores on Sintel, KITTI-2015 and Spring in a paper accepted at ECCV 2026.

The authors, Vladislav Bargatin, Alexander Yakovenko, Khaled Abud and Dmitriy Vatolin, describe the design in the arxiv abstract as "a hierarchical transformer built without any flow-specific components." Three attention variants carry the work: window attention for local processing, shifted-window attention for cross-window information exchange, and a global attention operating at reduced resolution.

The reported numbers are 0.68/1.48 EPE on Sintel Clean/Final, 3.23 Fl-all on KITTI-2015, and 3.192 1px on Spring, "while remaining memory efficient at 1080p inference." The abstract's own framing is that the standard components were not strictly required: "Despite the absence of standard inductive biases, FreeFlow achieves state-of-the-art results on major benchmarks."

The abstract does not publish parameter counts, training compute, or wall-clock inference speed, and it does not name the prior methods against which the state-of-the-art claim is made.