Priming Models: Teacher-Free One-Step Generation Hits FID 1.19 With 12x Less Compute
du_yilun · x · 2026-10-15
Jaeyeon Kim, with Sitan Chen, Yilun Du and others, introduce Priming Models, a new paradigm for training one-step generative models from scratch — no diffusion teacher, no flow map, no CFG, no JVP. The core recipe is [reconstruction pretraining → distributional post-training].
- 1.19 FID one-step on ImageNet 256×256 using substantially less compute: 12x less than Drifting-B, 6x less than Drifting-L
- Scales to text-to-image with GenEval 0.82
- Demonstrates world modeling on RT-1 / Bridge-V2
The authors believe the work will challenge how the field thinks about building one-step generative models.
More from Multimodal
- Kling 4.0 now fully available, showcased with 'King Card' short film — DavidmComfort · 2026-10-15
- Krea 3 image editing beta opens: 2k edits in under 8s, claims Nano Banana Pro-level quality — TitusTeatus · 2026-10-15
- Diffusers adds full Flux.3 Image Klein support: FP8/NVFP4 quantized inference and bbox layout control — RisingSayak · 2026-10-15
- xiaohu demos video model across styles with in-video text rendering — xiaohu · 2026-10-15
- Midjourney Style Code sref 2898442999 Yields Consistent Claymation-Style Anthropomorphic Animals — sergeantsref · 2026-10-15
- Kling 4.0 lip-sync demo handles even fast-paced rap — xiaohu · 2026-10-15