Improving Text-to-Image Generation with Intrinsic Self-Confidence Rewards

Researchers introduced ARC (Adaptive Rewarding by self-Confidence), a new framework for improving text-to-image generation models through self-confidence signals rather than external rewards. The method uses internal self-denoising probes to evaluate model accuracy and converts this into scalar rewards for unsupervised optimization, showing improvements in compositional generation and text-image alignment.

• →ARC framework eliminates the need for external reward supervision by using intrinsic self-confidence signals from the model itself.
• →The method evaluates model accuracy by testing how well it recovers injected noise under self-denoising probes.
• →ARC delivers consistent improvements in compositional generation, text rendering, and text-image alignment compared to baseline models.
• →The framework enables fully unsupervised optimization without requiring additional datasets, human annotators, or external reward models.
• →Combining ARC with external rewards provides complementary benefits while reducing reward hacking issues.