Curriculum Group Policy Optimization: Adaptive Sampling for Unleashing the Potential of Text-to-Image Generation

In three linesCGPO (Curriculum Group Policy Optimization) improves text-to-image model training via adaptive curriculum based on reward variance. Method prioritizes partially-mastered prompts (high variance) and balances categories through proportional fairness optimization. Gains validated on GenEval, T2I-CompBench++, DPG Bench.

Summary generated by Claude — human-verified