Training Design for Text-to-Image Models: Experimenting with Representation Alignment

AI Impact Summary

This document details an experimental logbook exploring training efficiency techniques for text-to-image models, specifically focusing on the PRX-1.2B model. The core approach involves aligning representations by adding a loss that directly supervises intermediate features using a frozen vision encoder. This technique, termed REPA, aims to accelerate convergence and improve training efficiency by leveraging the strong feature representations learned by pre-trained encoders. The experiment uses a baseline configuration with established metrics (FID, CMMD, DINO-MMD) to evaluate the impact of these interventions, providing a structured approach to optimizing model training.