Contribution
A fair comparison of latent spaces for generation, where “fair” means each space gets its own best timestep schedule, and a measurement of how much that tuning is worth.
Key insight
A latent space’s learnability cannot be read off one schedule. The optimal noise shift depends on the representation’s dimensionality and spectrum. Compared at a shared schedule, the space the schedule happens to suit wins.
Key results
- Best gFID: RAE 3.10, FLUX.2 3.70, SD 7.73, FLUX.1 10.13. LPIPS: FLUX.2 0.27, FLUX.1 0.34, SD 0.95, RAE 1.67
- At 300k steps, best-to-worst relative FID from the training shift alone: RAE 61.5%, FLUX.2 73.3%, SD 75.7%, FLUX.1 86.4%
- Logit-normal or plateau logit-normal beat shifted uniform in every latent. The optimal sampling shift is slightly above the optimal training shift
- Ranking flip: RAE loses to shifted FLUX.2 when RAE is unshifted, and wins when both are tuned
- REPA helps all four latents
Limitations
- One FID per configuration, one seed, ImageNet 256², DiT-XL. Random (not class-balanced) class sampling, so the numbers are not comparable to RAE’s paper
- The grid tops out at α = 6.93, which is where RAE’s optimum lands. The √(m/n) agreement is partly a grid-edge effect
- Table 2 (best parameters per step) is an interactive widget that did not render. The prose summary of it is what was read
- The authors’ own autoencoder, and the FLUX.2 model results are human preference only