zai-org/ImageReward

ImageReward is a pioneering general-purpose human preference reward model designed for text-to-image generation, trained on an extensive dataset of 137,000 expert comparisons. It significantly surpasses conventional text-image scoring methods like CLIP and Aesthetic in accurately reflecting human preferences. The project also introduces Reward Feedback Learning (ReFL), an innovative approach for directly optimizing text-to-image diffusion models using ImageReward, resulting in a 58.4% improvement in human evaluation over untuned versions. Available as a Python package, ImageReward empowers developers to both evaluate and fine-tune generative AI models, enhancing the alignment of synthesized images with human aesthetic and semantic expectations.