REPA

Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think

Sihyun Yu et al. · ICLR 2025 Oral

Read paper · arXiv:2410.06940

Align intermediate generative features with a frozen semantic teacher.

What you will implement

Patch-wise representation alignment and gradient boundaries.

• 40. REPA Representation Alignment

  Align noisy diffusion features patch by patch to frozen clean-image representations.

  Implement alignment and the combined objective on projected features; full DiT and teacher training are outside this module.