CodeGoat24/UnifiedReward-2.0-qwen-7b
UnifiedReward-2.0-qwen-7b by CodeGoat24 is a 7 billion parameter multimodal reward model built on Qwen/Qwen2.5-VL-7B-Instruct, designed for assessing image and video generation. It supports both pairwise ranking and pointwise scoring across dimensions like alignment, coherence, and style. This model is specifically optimized for vision model preference alignment, offering a unified approach to evaluating multimodal content.
Loading preview...
UnifiedReward-2.0-qwen-7b Overview
UnifiedReward-2.0-qwen-7b is a 7 billion parameter multimodal reward model developed by CodeGoat24. It is built upon the Qwen/Qwen2.5-VL-7B-Instruct architecture, making it capable of understanding and assessing both image and video generation. This model introduces a unified approach to evaluating visual content, distinguishing itself from other reward models by its comprehensive multimodal capabilities.
Key Capabilities
- Multimodal Assessment: Evaluates both image and video generation quality.
- Dual Scoring Methods: Supports both pairwise ranking and pointwise scoring for flexible evaluation.
- Detailed Evaluation Dimensions: Assesses content based on critical dimensions such as Alignment, Coherence/Physics, and Style.
- Vision Model Preference Alignment: Specifically designed to help align preferences for vision models, enhancing their output quality.
Good For
- Developers and researchers working on image and video generation models who need robust evaluation metrics.
- Aligning vision model outputs with human preferences.
- Comparative analysis of different generative models through pairwise ranking.
- Quantitative assessment of visual content quality using pointwise scoring.
For more technical details, refer to the accompanying paper and the project page.