Video World Models · 2026-10-15
Are We Overthinking One-Step Generation?
Woobin Park, Jaeyeon Kim, Yilun Du, Sitan Chen, Bohyung Han
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
The authors introduce Priming Models, a two-stage framework for one-step generators. Stage 1 trains with a reconstruction objective (e.g., pixel-space denoising at fixed noise). Stage 2 uses the pretrained network for feature-space distributional post-training. The authors report that this simple reconstruction stage alleviates the difficulty of distributional training and improves the training-compute vs. generation-quality trade-off, including a reported ImageNet FID of 1.19 for one-step generators.
Why it matters
One-step generative modeling enables single-network-evaluation sampling, but prior distributional one-step approaches required complex training recipes. The authors show that a simple reconstruction pretraining can make distributional post-training much more efficient, advancing performance under limited training compute and across multiple generative domains (class-conditional images, text-to-image, and world modeling).
Method
- Train a one-step generator in two stages: (1) reconstruction pretraining on corrupted-to-clean recovery (fixed noise level s=0.8 is sufficient in the main instantiation), then (2) distributional post-training by matching generated and real feature distributions with a differentiable feature-space distributional loss (e.g., OT/Sinkhorn-style).
- Use the pretrained network directly as a one-step generator mapping Gaussian noise to samples, then refine it by optimizing a feature-space objective with a frozen feature encoder and backpropagation through ϕ∘fθ.
- Evaluate on ImageNet class-conditional generation (pixel space), text-to-image generation (based on MiniT2I), and world modeling (autoregressive rollouts conditioned on actions), reporting compute/quality trade-offs.
Abstract (from arXiv)
Diffusion and flow matching are now dominant frameworks for generative modeling in continuous domains, but they typically require multiple network evaluations at inference. This has motivated extensive research on one-step generative modeling primarily built on diffusion and flow, though existing approaches often rely on complex training machinery. In this work, we introduce Priming Models, a simple two-stage framework for training one-step generators. Motivated by the difficulty of distributional training from scratch, Priming Models are first trained on a reconstruction objective, e.g., pixel-space denoising at a fixed noise level, before undergoing feature-space distributional post-training. Our central scientific finding is that this simple reconstruction stage is already enough to entirely alleviate the difficulty of distributional training. Empirically, across class-conditional image generation, text-to-image generation, and world modeling, Priming Models significantly advance the Pareto frontier of training compute vs. generation quality. On ImageNet at $256 \times 256$ resolution, we achieve a state-of-the-art FID of $1.19$ among one-step generators using approximately one-tenth the training compute of prior methods with comparable FID. We open-source our codebase at https://github.com/JaeyeonKim01/priming.
Related papers
- From Pixels, Without Pre-training: Joint Generative and Self-Supervised Representation Learning in One Model
- CAST: Causal Advantage-Structured Training with Spatially Grounded Compositional Rewards for Diffusion Models
- MoDDiT: Simple Token Routing in Diffusion Transformers
- Glance Forcing: Making Bidirectional Diffusion Autoregressive at a Glance
- LongLive-Plug: Once-for-All Distillation for Video Generation
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-14 · Indexed by Pub-AI: 2026-10-15 UTC
Are We Overthinking One-Step Generation?
Preprint posted: 2026-10-14 · Original record ↗
- arxiv
- 2610.20331
Explicit source versions (1)
- v1 · 2026-10-14
- Preprint available
Source date: 2026-10-14 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Diffusion and flow matching are now dominant frameworks for generative modeling in continuous domains, but they typically require multiple network evaluations at inference. This has motivated extensive research on one-step generative modeling primarily built on diffusion and flow, though existing approaches often rely on complex training machinery. In this work, we introduce Priming Models, a simple two-stage framework for training one-step generators. Motivated by the difficulty of distributional training from scratch, Priming Models are first trained on a reconstruction objective, e.g., pixel-space denoising at a fixed noise level, before undergoing feature-space distributional post-training. Our central scientific finding is that this simple reconstruction stage is already enough to entirely alleviate the difficulty of distributional training. Empirically, across class-conditional image generation, text-to-image generation, and world modeling, Priming Models significantly advance the Pareto frontier of training compute vs. generation quality. On ImageNet at $256 \times 256$ resolution, we achieve a state-of-the-art FID of $1.19$ among one-step generators using approximately one-tenth the training compute of prior methods with comparable FID. We open-source our codebase at https://github.com/JaeyeonKim01/priming.