Restoring Initial Noise Sensitivity in Text-to-Image Distillation via Geometric Alignment
Researchers have developed a new framework called Geometry-Aware Distillation (GAD) to improve text-to-image generation models. This method addresses the issue of lost sensitivity to initial noise in distilled models, which can hinder downstream control tasks. GAD works by aligning the local functional behavior of teacher and student models, specifically by matching Jacobian-vector products related to input noise. Experiments show GAD successfully restores noise sensitivity and enhances diversity while maintaining visual quality. AI
IMPACT Enhances control and diversity in text-to-image models by preserving crucial noise sensitivity during distillation.