REPA — Representation Alignment

Representation Alignment regularizes diffusion-transformer hidden states toward features from a pretrained visual encoder (e.g. DINOv2), giving a 17.5× SiT training speedup and FID 1.42 on ImageNet (ICLR 2025 oral). KAIST-led with Saining Xie as senior author; heavily adopted and extended (REPA-E, U-REPA, VA-VAE, JanusFlow, HunyuanVideo-Foley). Precursor to NYU's RAE line.