GT-SVJ: Generative-Transformer-Based Self-Supervised Video Judge For Efficient Video Reward Modeling

TL;DR

This paper introduces GT-SVJ, a novel approach that repurposes video generative models as reward models for evaluating video quality, leveraging their temporal modeling capabilities and contrastive training to outperform existing methods with fewer annotations.

Contribution

The paper presents a new method that transforms video generative models into temporally-aware reward models using energy-based reformulation and contrastive learning, reducing reliance on large annotated datasets.

Findings

Achieves state-of-the-art performance on GenAI-Bench and MonteBench.

Uses only 30K human annotations, significantly fewer than existing methods.

Effectively discriminates video quality through synthetic negative video generation.

Abstract

Aligning video generative models with human preferences remains challenging: current approaches rely on Vision-Language Models (VLMs) for reward modeling, but these models struggle to capture subtle temporal dynamics. We propose a fundamentally different approach: repurposing video generative models, which are inherently designed to model temporal structure, as reward models. We present the Generative-Transformer-based Self-Supervised Video Judge (\modelname), a novel evaluation model that transforms state-of-the-art video generation models into powerful temporally-aware reward models. Our key insight is that generative models can be reformulated as energy-based models (EBMs) that assign low energy to high-quality videos and high energy to degraded ones, enabling them to discriminate video quality with remarkable precision when trained via contrastive objectives. To prevent the model from exploiting superficial differences between real and generated videos, we design challenging synthetic negative videos through controlled latent-space perturbations: temporal slicing, feature swapping, and frame shuffling, which simulate realistic but subtle visual degradations. This forces the model to learn meaningful spatiotemporal features rather than trivial artifacts. \modelname achieves state-of-the-art performance on GenAI-Bench and MonteBench using only 30K human-annotations: $6\times$ to $65\times$ fewer than existing VLM-based approaches.