FreeFlow: A Bias-free Hierarchical Transformer for Optical Flow Estimation
Vladislav Bargatin, Alexander Yakovenko, Khaled Abud, Dmitriy Vatolin
AI Center, Lomonosov MSU · Lomonosov Moscow State University · MSU Institute for Artificial Intelligence
Published September 10, 2026 · on the feed September 11, 2026 · ▲ 34 · arXiv 2609.11486
The abstract
Optical flow methods typically rely on task-specific inductive biases, such as correlation volumes, feature warping, and iterative refinement, among others, to reach high accuracy. While effective, such biases constrain the model to predefined heuristics, which can limit its expressivity and lead to more complex pipelines and additional computational cost. We present FreeFlow, a hierarchical transformer built without any flow-specific components, using instead a single feed-forward encoder--decoder. FreeFlow combines three attention variants: window attention for local processing, shifted-window attention for cross-window information exchange, and a global attention operating at a reduced resolution. The resulting architecture scales naturally with model capacity, enabling a consistent accuracy gain from small to large variants. Despite the absence of standard inductive biases, FreeFlow achieves state-of-the-art results on major benchmarks, including Sintel (0.68/1.48 EPE on Clean/Final), KITTI-2015 (3.23 Fl-all), and Spring (3.192 1px), while remaining memory efficient at 1080p inference.
The Brief
2:02A short podcast summary of the paper. Written and voiced by AI.
Transcript
FreeFlow. A Bias-free Hierarchical Transformer for Optical Flow Estimation. By Vladislav Bargatin and three others, from the AI Center at Lomonosov Moscow State University, Lomonosov Moscow State University, and one other institution. Published on the tenth of September. Its code repository is worked out: about forty source files, with a readme and a licence.
Optical flow estimates how every pixel moves between two frames. To reach high accuracy, flow methods typically rely on inductive biases built specifically for the task: correlation volumes, feature warping, and iterative refinement, among others. These biases work, but they tie the model to predefined heuristics. That can limit how expressive the model is, and it tends to bring more complex pipelines and additional computational cost.
FreeFlow drops those components entirely. It is a hierarchical transformer with no flow-specific parts at all, just a single feed-forward encoder and decoder. The bet is that a general architecture, without heuristics designed for flow, can learn what the heuristics were providing.
FreeFlow combines three kinds of attention. Window attention handles local processing. Shifted-window attention lets information move between windows. And a global attention works at reduced resolution, so the whole frame can be seen at once. The resulting architecture scales naturally with model capacity, and accuracy rises consistently from the small variant to the large one.
Despite having none of the standard inductive biases, FreeFlow reaches state-of-the-art results on major benchmarks. On Sintel, its end-point error is zero point six eight on Clean and one point four eight on Final. On KITTI-2015, it scores three point two three on Fl-all. On Spring, it scores three point one nine two on the one-pixel metric. It also remains memory efficient when running inference at 1080p.
The result argues that the machinery optical flow has long relied on is not strictly required. A plain encoder and decoder, with attention at three scales, can match or pass the specialised designs, and it grows more accurate simply by being made larger.
Paper Quest · 7 chapters · about 41 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Optical flow has collected decades of clever parts. This quest asks whether it needs them.
You estimate optical flow: how every pixel moves between two frames. The best methods are packed with hand-designed parts, such as correlation volumes, feature warping and iterative refinement.
The road ahead
- 1 · The problem ~7 min Optical flow methods reach high accuracy with hand-designed parts, but those parts make pipelines complex, costly and hard to change.
- 2 · What came before ~3 min
- 3 · The key idea ~9 min Drop flow-specific modules and use one feed-forward transformer whose window, shifted-window and reduced-resolution global attention cover detail and large motions.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Tokens pass through window, shifted-window and reduced-resolution global attention blocks, then a simple head decodes the flow in one pass.
- 6 · The experiment ~6 min You’ll guess: Without correlation volumes, warping or iterative refinement, how did FreeFlow do on Sintel, KITTI-2015 and Spring?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.