Paper 2605.15141
Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation
- Published
- May 2026
- Research lab
- Independent
- Citations
- 11
- GitHub
- 905 stars
01 In brief
Summary
Causal Forcing++ is a scalable pipeline for real-time interactive video generation that distills bidirectional diffusion models into few-step autoregressive (AR) students.
It targets frame-wise autoregression with 1–2 sampling steps, a regime where existing initialization strategies fail: ODE distillation with a bidirectional teacher is architecturally misaligned, multi-step AR diffusion lacks few-step capability, and causal ODE distillation is too costly.
The key contribution is replacing causal ODE distillation with causal consistency distillation (causal CD), which learns the same AR-conditional flow map but uses a single online teacher ODE step between adjacent timesteps, avoiding precomputed PF-ODE trajectories.
This makes initialization more efficient and easier to optimize.
Experiments on Wan2.1-1.3B show Causal Forcing++ surpasses the SOTA 4-step chunk-wise Causal Forcing under frame-wise 2-step generation, improving VBench Total by 0.1, VBench Quality by 0.3, and VisionReward by 0.335, while reducing first-frame latency by 50% and Stage 2 training cost by ~4×.
The method also extends to action-conditioned world models in the spirit of Genie3.
02 From the paper
Abstract
Real-time interactive video generation requires low-latency, streaming, and controllable rollout. Existing autoregressive (AR) diffusion distillation methods have achieved strong results in the chunk-wise 4-step regime by distilling bidirectional base models into few-step AR students, but they remain limited by coarse response granularity and non-negligible sampling latency. In this paper, we study a more aggressive setting: frame-wise autoregression with only 1--2 sampling steps. In this regime, we identify the initialization of a few-step AR student as the key bottleneck: existing strategies are either target-misaligned, incapable of few-step generation, or too costly to scale. We propose \textbf{Causal Forcing++}, a principled and scalable pipeline that uses \emph{causal consistency distillation} (causal CD) for few-step AR initialization. The core idea is that causal CD learns the same AR-conditional flow map as causal ODE distillation, but obtains supervision from a single online teacher ODE step between adjacent timesteps, avoiding the need to precompute and store full PF-ODE trajectories. This makes the initialization both more efficient and easier to optimize. The resulting pipeline, \ours, surpasses the SOTA 4-step chunk-wise Causal Forcing under the \textit{\textbf{frame-wise 2-step setting}} by 0.1 in VBench Total, 0.3 in VBench Quality, and 0.335 in VisionReward, while reducing first-frame latency by 50\% and Stage 2 training cost by $\sim$$4\times$. We further extend the pipeline to action-conditioned world model generation in the spirit of Genie3. Project Page: https://github.com/thu-ml/Causal-Forcing and https://github.com/shengshu-ai/minWM .