MJ-Video: Benchmarking and Rewarding Video Generation with Fine-Grained Video Preference

Recent advancements in video generation have significantly improved the ability to synthesize videos from text instructions. However, existing models still struggle with key challenges such as instruction misalignment, content hallucination, safety concerns, and generation bias. To address these limitations, we introduce MJ-B ENCH -V IDEO , a large-scale video preference benchmark designed to evaluate video generation across five critical aspects: Alignment, Safety, Fineness, Coher-ence & Consistency, and Bias & Fairness . This benchmark further incorporates 28 fine-grained criteria to provide a comprehensive evaluation of video preference. Building upon this dataset, we propose MJ-V IDEO , a Mixture-of-Experts (MoE)- based video reward model designed to deliver fine-grained reward. MJ-V IDEO can dynamically select relevant experts to accurately judge the preference based on the input text-video pair. This architecture enables more precise and adaptable preference judgments. Through extensive benchmarking on MJ-B ENCH -V IDEO , we analyze the limitations of existing video reward models and demonstrate the superior performance of MJ-V IDEO in video preference assessment, achieving 17.58% and 15.87% improvements in overall and fine-grained preference judgments, respectively. Additionally, MJ-V IDEO is able to improve the alignment performance in video generation via preference fine-tuning. Warning: this paper contains content that may be inappropriate or offensive.