프로그래밍선

RAE(Representation Autoencoder) 요약Encoder: frozen pretrained representation encoder (예: DINO / SigLIP / MAE)Decoder: 그 representation을 받아서 픽셀을 복원하도록 새로 학습한(lightweight) decoderfrozen representation encoder가 만든 feature/latent를 받아서 픽셀로 복원하는 reconstruction decoder.DiT-DH는 DiT 위에 붙는 추가 headshallow-but-wide diffusion transformer head를 기존 DiT 백본(backbone) 뒤에 덧붙이는 구조. 목적은 고차원 RAE latent에서 생기는 차원 병목을 완화..

REPA 를 뛰어넘을 논문이 나왔다! REG ! Representation Entanglement for Generation:Training Diffusion Transformers Is Much Easier Than You Think[Submitted on 2 Jul 2025]https://arxiv.org/abs/2507.01467 Representation Entanglement for Generation:Training Diffusion Transformers Is Much Easier Than You ThinkREPA and its variants effectively mitigate training challenges in diffusion models by incorporating exter..