Priming Models: one-step generation hits 1.19 FID with 12x less compute, no diffusion teacher
LucaAmb · x · 2026-10-20
Jaeyeon Kim et al. introduce Priming Models, a new paradigm for training one-step generative models from scratch with no diffusion teacher, flow map, CFG, or JVP. Core idea: reconstruction pretraining provides a good initialization, then distributional post-training does the heavy lifting. It achieves 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, and shows world modeling on RT-1/Bridge-V2. Commenters note pretraining may mainly need to find a good initialization, though scaling may share flaws with drifting models.
More from Multimodal
- Creator generates tense cliff-road car clip in 30 seconds with Seedance 2.5, prompt included — LudovicCreator · 2026-10-20
- MOSS-Image ecosystem roundup: layout editors, line-art colorizer, color-depth control — optimisticalish · 2026-10-20
- LLM-composed piano beats text-to-music models in 4-way AI composition comparison — UnfortunateHurricane · 2026-10-20
- Kling AI 4.0 debuts with 30-second clips, stereo audio, but doesn't outpace Wan 3.0 and Seedance 2.5 — The Decoder · 2026-10-20
- Runway Labs unveils Continuum, an OS-like research app for real-time video interfaces — jongranskog · 2026-10-20
- Fish Drama-3 lands in HF speech-to-speech: voice cloning with inline whisper, laugh and sigh cues — andimarafioti · 2026-10-20