Human detectors are surprisingly powerful reward models

TL;DR

This paper introduces HuDA, a simple reward model leveraging off-the-shelf human detection and motion alignment to significantly improve the realism of generated videos with complex human motions, outperforming existing models.

Contribution

HuDA is a novel, training-free reward model that enhances human motion realism in video generation, applicable to various models and beyond humans.

Findings

HuDA outperforms specialized models with manual annotations.

Using HuDA in GRPO improves complex human motion generation.

HuDA enhances video quality for animals and human-object interactions.

Abstract

Video generation models have recently achieved impressive visual fidelity and temporal coherence. Yet, they continue to struggle with complex, non-rigid motions, especially when synthesizing humans performing dynamic actions such as sports, dance, etc. Generated videos often exhibit missing or extra limbs, distorted poses, or physically implausible actions. In this work, we propose a remarkably simple reward model, HuDA, to quantify and improve the human motion in generated videos. HuDA integrates human detection confidence for appearance quality, and a temporal prompt alignment score to capture motion realism. We show this simple reward function that leverages off-the-shelf models without any additional training, outperforms specialized models finetuned with manually annotated data. Using HuDA for Group Reward Policy Optimization (GRPO) post-training of video models, we significantly enhance video generation, especially when generating complex human motions, outperforming state-of-the-art models like Wan 2.1, with win-rate of 73%. Finally, we demonstrate that HuDA improves generation quality beyond just humans, for instance, significantly improving generation of animal videos and human-object interactions.