AutoRubric-T2I: Robust Rule-Based Reward Model for Text-to-Image Alignment

AutoRubric-T2I replaces Bradley-Terry preference models for T2I reward modeling with VLM-generated rule-based rubrics, reducing reliance on large human preference datasets. The approach improves transparency and adaptability over opaque BT models while offering finer-grained scoring than heuristic VLM judges. Practitioners building RLHF pipelines for image generation should evaluate this as a cheaper, more interpretable alternative.