Abstract
Recent advances in human preference alignment have significantly enhanced multimodal generation and understanding. A key approach is training reward models to guide preference optimization. However, existing models are often task-specific, limiting their adaptability across diverse visual applications. We also argue that jointly learning to assess multiple tasks may foster a synergistic effect, where improved image understanding enhances image generation assessment, and refined image evaluation benefits video assessment through better frame analysis. To this end, this paper proposes UnifiedReward, the first unified reward model for multimodal understanding and generation assessment, enabling both pairwise ranking and pointwise scoring, which can be employed for vision model preference alignment. Specifically, (1) we first develop UnifiedReward on our constructed large-scale human preference dataset, including both image and video generation/understanding tasks. (2) Then, it is utilized to automatically construct high-quality preference pair data based on the vision models, fine-gradually filtering their outputs through pair ranking and point sifting. (3) Finally, these data are used for their preference alignment through Direct Preference Optimization (DPO). Experimental results demonstrate that joint learning to assess diverse visual tasks can lead to substantial mutual benefits and we apply our pipeline to both image and video understanding/generation tasks, significantly improving the performance in each domain.
Introduction
In the development of Multimodal Large Language Models (MLLMs) and Diffusion models, "alignment" is the critical process of ensuring model outputs resonate with human preferences. This typically necessitates a reward model—an evaluative component that distinguishes between high-quality and low-quality outputs.
However, current reward models face two primary limitations:
- Task-Specificity: Most existing models are specialized for a single domain. For instance, a model designed to evaluate image generation (e.g., PickScore) cannot effectively judge video understanding (e.g., LLaVA-Critic). This lack of versatility restricts their application across the broader multimodal landscape.
- Untapped Synergies: Visual tasks are inherently interconnected. A model capable of deep image understanding should, in theory, be better equipped to judge the realism of a generated image. Current research often overlooks the potential benefits of cross-task learning.
Methods
The authors propose UnifiedReward, built upon the LLaVA-OneVision-7B architecture. The framework employs a comprehensive three-stage pipeline:
- Unified Reward Model Training: The base model is fine-tuned on a newly curated dataset of 236,000 samples. The model is uniquely trained to perform both pairwise ranking (comparing two outputs) and pointwise scoring (assigning an absolute quality metric).
- Two-Stage Preference Data Construction: UnifiedReward is used to generate high-quality alignment data through:
- Pair Ranking: Organizing model outputs into ranked groups.
- Point Sifting: Utilizing pointwise scores to filter out "noisy" pairs where the quality difference is negligible, ensuring a clear preference signal.
- Model Alignment: The refined data is used to fine-tune vision models (such as SDXL or LLaVA) via Direct Preference Optimization (DPO).
Datasets
To support this unified approach, the researchers aggregated and standardized several existing sources into a 236K sample preference dataset:
- Image Generation: EvalMuse, Human Preference Dataset (HPD), and Open-Image-Preferences (OIP).
- Image Understanding: LLaVA-Critic-113K.
- Video Generation: VideoDPO, LiFT-HRA, and VideoFeedback.
- Video Understanding: ShareGPTVideo-DPO.
Key Insights
- The Synergistic Effect: Joint training across understanding and generation tasks proves highly effective. For example, incorporating image understanding tasks improved the model's accuracy in evaluating image generation compared to models trained solely on generation data.
- Point Sifting for Quality: Moving beyond simple ranking, "point sifting" filters for pairs with significant score gaps. This provides a cleaner, more robust signal for DPO, resulting in superior model alignment.
- Cross-Modality Transfer: Proficiency in image evaluation naturally extends to video. Enhanced frame-level analysis (derived from image tasks) leads to more accurate assessments of temporal consistency in video generation.
Key Results
UnifiedReward consistently outperforms specialized state-of-the-art reward models across multiple benchmarks:
- Image Understanding: On VLRewardBench, UnifiedReward achieved 66.1% accuracy, surpassing LLaVA-Critic (46.9%) and even GPT-4o (65.8%).
- Image Generation: On GenAI-Bench, it reached a Tau score of 54.8, exceeding PickScore (53.2) and HPSv2 (51.6).
- Video Generation: It achieved a Tau of 60.7 on GenAI-Bench (Video), a significant improvement over VideoScore (46.2).
- Scalability: Testing with larger backbones (e.g., Qwen2.5-VL-32B) showed that performance scales with model size, reaching 81.5% accuracy on VLRewardBench.
Limitations and Future Work
While UnifiedReward represents a significant step forward, several areas remain for exploration:
- Data Imbalance: Video generation preference data is currently less abundant than image data. Future work will focus on expanding high-quality video datasets.
- Computational Cost: Operating a 7B or 32B parameter model as a reward judge is more resource-intensive than using smaller, task-specific scorers.
- Expanded Modalities: The authors suggest extending the unified framework to include 3D or audio-visual tasks and exploring "Online" RLHF applications.
Summary updated