Researchers introduced UniEvo-VL, a self-improvement method for multimodal models. It uses a model’s own critiques as privileged information during training. The model acts as both teacher and student with different contexts. Training minimizes divergence between its own sampling trajectories. Tests on Qwen-image-2512 showed gains on GenEval and GenEval2 Soft-TIFA. Stronger external critics raised the self-improvement ceiling further. Mixed text-rendering results suggest uneven improvements across tasks.
What changed
Multimodal models relied on separate, often larger teacher models for self-improvement.
What it unlocks
A single multimodal model can act as both teacher and student to improve its own image generation.
- 0.747 to 0.808 on GenEval
- 32.97 to 35.53 on GenEval2 Soft-TIFA
What you need to act on it
- a multimodal model with self-critique capability
- test-time compute access
Sources