JLT: Clean-Latent Prediction in Latent Diffusion Transformers

Funing Fu, Tenghui Wang, Junyong Cen, Qichao Zhu, Guanyu Zhou

arXiv:2605.27102 · 2026-05-27 공개 · arXiv · PDF

transformer flow-matching latent-space image-net latent-diffusion fid-metric clean-latent-prediction flux-vae

Abstract

Flow matching with clean-data prediction has shown that regressing the clean point can exploit low-dimensional structure more effectively than predicting an ambient noised quantity. We ask whether this principle remains useful after images are mapped into a learned latent space, where compression has already removed much of the raw pixel variability. We introduce JLT, a 130M latent diffusion Transformer over frozen FLUX.2 VAE codes, and compare clean-latent prediction with a matched velocity-prediction DiT under the same representation, backbone, and training settings. Although the three variables x, epsilon, and v are linearly convertible for a fixed corruption time, a local Gaussian analysis shows that velocity regression inherits an isotropic target-covariance floor and amplifies low-variance latent directions, while clean prediction damps them. On ImageNet 256 x 256, JLT-B/1 obtains FID-50K 2.50 with classifier-free guidance, with a large matched-target gap over velocity prediction. These results suggest that prediction targets in latent diffusion are representation-dependent geometric choices, rather than interchangeable algebraic parameterizations.

한국어 요약

📋 한 줄 요약

**[Latent Diffusion / Clean Prediction]** JLT가 FLUX.2 VAE 잠재공간에서 clean-latent 예측 DiT 검증, velocity 예측 대비 isotropic covariance floor 회피·ImageNet 256² FID 2.50으로 매치된 velocity baseline 대비 큰 격차.

🎯 핵심 기여도

💡 핵심 아이디어

Latent diffusion에서 예측 target(x, ε, v) 선택은 단순 algebraic parameterization 교체가 아니라 representation-dependent geometric choice이며, learned latent space의 variance 구조에 대해 clean prediction은 low-variance direction을 dampen하고 velocity prediction은 amplify해 본질적으로 다른 행동을 보인다.

🔬 기술적 접근법

📊 주요 결과

💭 의의 및 한계

**의의**: Latent diffusion의 prediction target 선택이 단순 algebraic 교환이 아니라 representation-dependent geometric 결정임을 이론·실험으로 명확화, FLUX.2 VAE의 latent space에서 clean prediction 우월성 정량 입증, 130M scale에서 SOTA에 근접 FID. **한계**: 130M·ImageNet 단일 평가, frozen FLUX.2 VAE에 종속, 매우 큰 모델 scale에서 동작 추가 검증, text-to-image 등 conditional generation 일반화는 후속.

🚀 실용적 활용