PRISM: Prompt Refinement via Image-grounded Self-rewarding Mechanism for Text-to-Image Generation

Researchers introduced PRISM, an image-grounded self-rewarding framework for refining prompts in text-to-image generation.

The system interprets generated images with structured visual feedback and scores them on semantic consistency, aesthetics, and human preference.

It uses that signal to improve prompt policies, yielding better results than text-only prompt optimization methods.

Experiments showed stronger image quality and semantic alignment, while also producing reusable and interpretable refinement feedback.