Alignment-Guided Score Matching for Text-to-Image Alignment in Diffusion Models

Researchers introduce Alignment-Guided Score Matching, a lightweight post-training method that improves text-to-image alignment in diffusion models without external rewards.

The approach injects alignment guidance into the score-matching objective and refines soft text tokens, making it compatible with major backbones like Stable Diffusion 1.5, SDXL, and SD3.

It matches SoftREPA overall while improving harder failure cases such as repetition and counting, including a reported 35%+ gain in GenEval counting accuracy.