FreeFlow: A Bias-free Hierarchical Transformer for Optical Flow Estimation

Hugging Face Daily Papers Papers
optical-flow transformer computer-vision deep-learning hierarchical bias-free

Summary

FreeFlow is a hierarchical transformer for optical flow estimation that eliminates task-specific inductive biases and achieves state-of-the-art accuracy on benchmarks like Sintel and KITTI-2015.

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.
Original Article
View Cached Full Text

Cached at: 09/11/26, 10:16 AM

Paper page - FreeFlow: A Bias-free Hierarchical Transformer for Optical Flow Estimation

Source: https://huggingface.co/papers/2609.11486

Abstract

FreeFlow is a hierarchical transformer for optical flow that eliminates task-specific inductive biases and achieves state-of-the-art accuracy using window, shifted-window, and global attention.

Optical flowmethods typically rely on task-specific inductive biases, such ascorrelation volumes,feature warping, anditerative 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, ahierarchical transformerbuilt without any flow-specific components, using instead a single feed-forward encoder--decoder. FreeFlow combines three attention variants:window attentionfor local processing,shifted-window attentionfor cross-window information exchange, and aglobal attentionoperating 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, includingSintel(0.68/1.48 EPE on Clean/Final),KITTI-2015(3.23 Fl-all), andSpring(3.192 1px), while remaining memory efficient at 1080p inference.

View arXiv pageView PDFProject pageGitHub2Add to collection

Get this paper in your agent:

hf papers read 2609\.11486

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper18

#### a-yakovenko/FreeFlow-S-pretrain Image-to-Image• 26.8M• Updatedabout 1 hour ago #### a-yakovenko/optical-flow-FreeFlow-S-TaTSKH Image-to-Image• 34.6M• Updatedabout 1 hour ago #### a-yakovenko/optical-flow-FreeFlow-S-TaTSKH-HQ Image-to-Image• 34.6M• Updatedabout 1 hour ago #### a-yakovenko/optical-flow-FreeFlow-S-sintel Image-to-Image• 34.6M• Updatedabout 1 hour ago Browse 18 models citing this paper## Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2609.11486 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2609.11486 in a Space README.md to link it from this page.

Collections including this paper3

Similar Articles

Flow Map Learning via Nongradient Vector Flow

arXiv cs.LG

This paper introduces SGFlow, a method for learning flow maps for diffusion models that avoids invertibility constraints and backpropagation through model iterations, achieving competitive FID scores on CIFAR with a proven stationary-point guarantee.

ChangeFlow -- Latent Rectified Flow for Change Detection in Remote Sensing

Hugging Face Daily Papers

ChangeFlow presents a generative framework for remote sensing change detection that synthesizes change masks in latent space using rectified flow, achieving improved accuracy and robustness through sampling-based prediction ensembling, with an average F1 of 80.4% across four benchmarks.

MeshFlow: Mesh Generation with Equivariant Flow Matching

Hugging Face Daily Papers

MeshFlow introduces an equivariant optimal-transport flow matching model for direct triangle mesh generation, achieving state-of-the-art quality while providing approximately 18x inference speedup over autoregressive methods.