Open-DDVM: A Reproduction and Extension of Diffusion Model for Optical Flow Estimation

TL;DR

This paper reproduces and extends the open-source diffusion model DDVM for optical flow estimation, demonstrating comparable performance with publicly available data and resources, and providing insights into key design choices.

Contribution

It presents the first open-source implementation of DDVM, reproduces its results, and analyzes important design choices for optical flow estimation.

Findings

Reproduced DDVM achieves comparable performance to the original.

Identified key design choices impacting model performance.

Provided open-source code and trained model for community use.

Abstract

Recently, Google proposes DDVM which for the first time demonstrates that a general diffusion model for image-to-image translation task works impressively well on optical flow estimation task without any specific designs like RAFT. However, DDVM is still a closed-source model with the expensive and private Palette-style pretraining. In this technical report, we present the first open-source DDVM by reproducing it. We study several design choices and find those important ones. By training on 40k public data with 4 GPUs, our reproduction achieves comparable performance to the closed-source DDVM. The code and model have been released in https://github.com/DQiaole/FlowDiffusion_pytorch.