Human detectors are surprisingly powerful reward models

TL;DR

HuDA introduces a zero-shot reward model for assessing human appearance and motion in generated videos by combining a framewise human-detection score with a temporal prompt-alignment score. Implemented as a simple linear combination, HuDA is used to post-train video diffusion models via Group Reward Policy Optimization (GRPO), achieving state-of-the-art improvements on complex human actions (e.g., 73% win-rate on hard prompts against Wan 2.1). The approach relies solely on off-the-shelf components (ViTDet, Llama, BLIP) and requires no manual annotations, yet correlates strongly with human judgments and preserves prompt faithfulness. Importantly, HuDA generalizes beyond humans to improve animal videos and human–object interactions, indicating broad applicability for realistic deformable-object video generation.

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.

Figures (12)

• Figure 1: HuDA is a reward model to quantify the human appearance and motions in generated videos. HuDA detects extra or missing body parts, implausible body pose, and misalignment with the prompt. Training a video diffusion model with HuDA using GRPO results in improved human appearance and motion in generated videos (bottom) compared to state-of-the-art baseline (top). Videos corresponding to all figures in the paper are provided in the \ref{['sec:appdx:videos']}.
• Figure 2: SORA-2 and Veo-3 generations for the prompts "@rgirdhar (author) doing a backflip," and "A person doing a backflip in a gym." Even state-of-the-art proprietary models struggle to generate complex human actions. Full video in \ref{['sec:appdx:videos']}.
• Figure 3: Our reward model HuDA quantifies the human detection score and temporal prompt alignment scores. The first term is the confidence level of the human detector in the worst performing window $W$ shown in purple band. Next, temporal prompt alignment score is the average of similarity between phase descriptions $T_i$ and uniformly sampled frames. Overall reward is a weighted average of human detection score and temporal prompt alignment.
• Figure 4: Video generation win-rate between HuDA (depicted by colored bars) and previous state-of-the-art baselines ($100 -$HuDA's win rate). Our method is consistently preferred by human annotators across all settings, with biggest gains in the hard category (avg win-rate $73\%$ vs $63\%$ for easy, across all baselines).
• Figure 5: Qualitative comparison of videos generated by our method (bottom) with Wan 2.1 14B wan (top). Wan 2.1 generates distorted poses (left), two-headed person (middle), and inconsistent legs (right), whereas Wan 2.1 trained with HuDA generates realistic humans.