Announcement
Priming Models propose training one-step generative models from scratch
The researchers say the method pairs reconstruction pretraining with distributional post-training, without relying on a diffusion teacher.
TLDR
The researchers introducing Priming Models say their approach trains one-step generative models from scratch, without a diffusion teacher or flow map. They report a 1.19 FID score on ImageNet 256×256, using 12 times less compute than Drifting-B and six times less than Drifting-L. They also cite text-to-image generation with a 0.82 GenEval score and world modeling on RT-1 and Bridge-V2.
Combined views
75.3K
4 Sources, first seen ago
