UniEvo-VL: Self-Distillation Training for Multimodal Model Self-Improvement

UniEvo-VL is a self-distillation framework that enables multimodal models to improve themselves by learning from their own self-critiques during test-time, acting as both teacher and student with different contexts. The approach significantly improves image generation performance—boosting scores from 0.747 to 0.808 on GenEval and from 32.97 to 35.53 on GenEval2 Soft-TIFA—without requiring external supervision or separate teacher models.