SOTA-346: Retune the training timestep shift for each autoencoder latent space, and never rank latent spaces for generation under one shared schedule

Source

Black Forest Labs (2025), LIT-572. Read as NOTE-311.

The practice

• Changing the autoencoder changes the noise schedule you need. The shift that suits a 16-channel SD latent is not the one that suits a 128- or 768-channel latent. In the report the best training shift ranged from α = 1 (SD) to at least 6.93 (RAE), and choosing it badly cost 61–86% relative FID
• Sweep the training shift first, then the sampling shift. The sampling shift mattered less (4.5–38.7%), and its optimum sat slightly above the training shift’s
• Use a logit-normal training distribution (as in SOTA-266). It, or its plateau variant, beat shifted-uniform sampling in every latent
• When comparing latent spaces, tune each. Otherwise the comparison measures fit to the schedule. The report’s ranking of RAE against FLUX.2 reverses between untuned and tuned

Conditions

• Single runs on ImageNet 256² with a DiT-XL. Text-to-image at scale is not tested
• The report’s √(m/n) rule for predicting the shift is not established by its own data (see LIT-572). Sweep it rather than computing it. The resolution form of the same argument is SOTA-263