Priming Models: 1.19 FID one-step generation without diffusion teacher, 12x less compute

LucaAmb · x · 2026-11-04

Priming Models: a teacher-free one-step generation paradigm

Jaeyeon Kim and collaborators (including Yilun Du, Bohyung Han) introduce Priming Models, a new paradigm for training one-step generative models from scratch: no diffusion teacher, flow map, CFG, or JVP.
Core message: reconstruction pretraining → distributional post-training. Stage 1 trains a denoiser at one fixed interpolation time; Stage 2 (much more expensive) turns it into a one-step generator.
Results: 1.19 FID one-step on ImageNet 256×256 with substantially less compute — 12x less than Drifting-B, 6x less than Drifting-L; scales to text-to-image (GenEval 0.82) and world modeling on RT-1 / Bridge-V2.
Moritz Piening then ran follow-up experiments on CIFAR-10/MNIST/Fashion-MNIST from a flow-matching and optimal-transport perspective, finding Stage-1 init highly sensitive while optimal transport couplings help somewhat; full write-up on his blog.

More from Multimodal