MoDiPO: text-to-motion alignment via AI-feedback-driven Direct Preference Optimization

TL;DR

MoDiPO introduces an AI-feedback-driven approach to align text-to-motion diffusion models, improving realism and controllability while reducing the need for human preference data.

Contribution

The paper presents MoDiPO, a novel method using AI feedback for Direct Preference Optimization to enhance text-to-motion model alignment, with a new motion-preference dataset.

Findings

Significantly improves motion realism as measured by FID

Maintains comparable RPrecision and multi-modality performance

Reduces reliance on human preference data in DPO

Abstract

Diffusion Models have revolutionized the field of human motion generation by offering exceptional generation quality and fine-grained controllability through natural language conditioning. Their inherent stochasticity, that is the ability to generate various outputs from a single input, is key to their success. However, this diversity should not be unrestricted, as it may lead to unlikely generations. Instead, it should be confined within the boundaries of text-aligned and realistic generations. To address this issue, we propose MoDiPO (Motion Diffusion DPO), a novel methodology that leverages Direct Preference Optimization (DPO) to align text-to-motion models. We streamline the laborious and expensive process of gathering human preferences needed in DPO by leveraging AI feedback instead. This enables us to experiment with novel DPO strategies, using both online and offline generated motion-preference pairs. To foster future research we contribute with a motion-preference dataset which we dub Pick-a-Move. We demonstrate, both qualitatively and quantitatively, that our proposed method yields significantly more realistic motions. In particular, MoDiPO substantially improves Frechet Inception Distance (FID) while retaining the same RPrecision and Multi-Modality performances.