The authors first isolate components of REPA-E to determine what drives its diffusability improvements.
Robust prediction principle yields better latent spaces for diffusion models
RoAR-VAE shows that recovering clean targets from corrupted latents improves both reconstruction fidelity and downstream generation quality.
Big Tech
Yang Yue · Fangyun Wei · Yitong Wang · Jinjing Zhao · Henry Zheng · Dong Chen · +1 more
Tsinghua University · Microsoft Research
Research Digest··2 min read
The authors revisit REPA-E, a prior approach for improving latent diffusability, and find that removing the diffusion objective while retaining semantic prediction from corrupted latents gives comparable results.
§