Priming Models introduced as a way to train one-step generative models from scratch

The team says Priming Models pair reconstruction pretraining with distributional post-training, without a diffusion teacher or flow map. It reports 1.19 FID in one step on ImageNet 256×256 with 12X less compute than Drifting-B and 6X less than Drifting-L. The team also cites a 0.82 GenEval score for text-to-image generation and world modeling on RT-1 and Bridge-V2. A commenter calls the results interesting but suggests scaling may share drifting models’ flaws.