Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
This paper investigates the impact of different prediction targets (clean data vs. velocity) within latent diffusion models using a 130M parameter latent diffusion Transformer (JLT) built on frozen FLUX.2 VAE codes. They find that predicting the clean latent representation outperforms velocity prediction, achieving an FID-50K of 2.50 on ImageNet 256x256. A local Gaussian analysis reveals that velocity regression introduces an isotropic target-covariance floor and amplifies low-variance latent directions, while clean prediction damps them, explaining the performance difference.
Clean-data prediction still beats velocity prediction even *after* images are mapped into a learned latent space, suggesting that the choice of prediction target in latent diffusion is a crucial geometric decision, not just an algebraic convenience.
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.