Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation

Min Zhao; Hongzhou Zhu; Kaiwen Zheng; Zihan Zhou; Bokai Yan; Xinyuan Li; Xiao Yang; Chongxuan Li; Jun Zhu

Source-linked AI summary

Causal Forcing++: Scalable Few-Step Autoregressive Diffusion Distillation for Real-Time Interactive Video Generation

Min Zhao, Hongzhou Zhu, Kaiwen Zheng, Zihan Zhou, Bokai Yan, Xinyuan Li, Xiao Yang, Chongxuan Li, Jun Zhu

arXiv:2605.15141v3May 14, 2026cs.CV

TL;DR

Real-time interactive video generation needs lower-latency, finer-grained autoregressive diffusion, but existing methods largely target chunk-wise four-step generation and costly initialization. Causal Forcing++ uses causal consistency distillation for frame-wise few-step initialization, achieving stronger overall performance than prior methods with 50% lower first-frame latency.

Problem
Existing methods remain insufficiently validated for frame-wise 1–2-step generation, while their full-trajectory initialization is costly to scale.
Method
Causal Forcing++ initializes few-step autoregressive students with causal consistency distillation using a single online teacher ODE step instead of stored full trajectories.
Results
Causal Forcing++ achieves the best overall performance among existing methods in frame-wise 2-step generation while reducing first-frame latency by 50%.
Takeaways & Limitations
Causal consistency distillation provides an efficient alternative to causal ODE initialization for low-latency frame-wise video generation.
Takeaways & Limitations
Causal ODE initialization remains difficult to scale because it requires generating, storing, and regenerating full multi-step teacher trajectories.

Abstract

from arXiv · show

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 .

1. Introduction

Causal Forcing++ addresses few-step frame-wise autoregressive video generation by using causal consistency distillation to initialize AR students efficiently and with causal alignment. On Wan2.1-1.3B, it improves quality and reward while reducing first-frame latency and Stage 2 training cost.

Problem: Few-step AR student initialization is the key bottleneck for frame-wise 1–2 step generation before asymmetric DMD.Existing initialization strategies are target-misaligned, lack few-step capability, or are costly to scale.
Method: Causal Forcing++ uses causal consistency distillation to learn the AR-conditional flow map for few-step AR student initialization.Causal consistency distillation targets the same flow map as causal ODE distillation while using supervision from an online teacher ODE step between adjacent timesteps.
Results: ~4× lower Stage 2 training cost accompanies Causal Forcing++’s frame-wise 2-step performance gains.Experiments on Wan2.1-1.3B validate the pipeline in the aggressive low-latency regime.

2. Background

Diffusion and flow-matching models generate samples by learning score- or velocity-related information and solving the PF-ODE. Autoregressive diffusion improves video interactivity, while distillation methods address its remaining multi-step latency through staged training pipelines.

Generative modeling through diffusion: Diffusion models perturb data with a noise schedule and train a velocity model using flow matching.The perturbation is x_t = α(t)x_0 + σ(t)ϵ, while flow matching uses α(t) = 1 − t and velocity regression.
Generative modeling through diffusion: Solving the PF-ODE with an optimal velocity model generates samples from the data distribution.The PF-ODE evolves x_t from t = 1 to 0 using the learned velocity field.
Autoregressive diffusion for video generation: Bidirectional video diffusion generates entire videos in one pass but incurs high first-frame latency and requires complete conditioning signals upfront.These properties make bidirectional generation non-interactive for user-driven applications.
Autoregressive diffusion for video generation: Autoregressive diffusion performs autoregression across frames or chunks and diffusion within each unit, using causal attention and KV-cache self-rollouts.This design enables interactivity compared with bidirectional video generation.
AR diffusion distillation: CausVid, Self Forcing, and Causal Forcing progressively refine AR diffusion distillation, while the resulting pipeline uses teacher forcing, causal ODE initialization, and student self-rollout asymmetric DMD.AR diffusion distillation is motivated by the latency of multi-step generation; Causal Forcing addresses the architectural mismatch in bidirectional-teacher ODE initialization.

3. Method

Causal Forcing++ addresses the lack of an AR, few-step, scalable initialization for aggressive frame-wise diffusion generation by replacing causal ODE distillation with causal consistency distillation. The method shares causal ODE distillation’s target while using local teacher supervision, improving efficiency, optimization, and autoregressive robustness.

Initialization Challenges: Existing initializations fail in aggressive low-latency regimes because they are architecturally misaligned, lack few-step capability, or are too costly to scale.Bidirectional-teacher ODE distillation can collapse, direct multi-step AR initialization weakens toward frame-wise 1-step generation, and causal ODE distillation is difficult to scale.
Causal Consistency Distillation: Causal consistency distillation learns the same AR flow map as causal ODE distillation, making it a principled few-step AR initialization.The two methods differ in supervision: ODE distillation uses large jumps on pre-generated trajectories, whereas causal CD enforces consistency locally between adjacent timesteps on real data.
Causal Consistency Distillation: One teacher ODE step per iteration replaces full offline PF-ODE trajectories, eliminating trajectory storage and regeneration while reducing the optimization gap to Δt.Causal ODE distillation may require pre-generating and storing trajectories of 48 steps, whereas causal CD uses adjacent timestep pairs.
Causal Consistency Distillation: Causal CD matches or surpasses causal ODE distillation on VBench at both Stage 2 and after asymmetric DMD across chunk-wise and frame-wise configurations.The local-pairing scheme also yields a stronger initialization through an easier per-step optimization target.
Causal Forcing++: Causal Forcing++ retains teacher-forcing AR training and asymmetric DMD with self-rollout, replacing Causal Forcing’s Stage 2 with the causal CD objective.This composition targets improved AR diffusion distillation while preserving the surrounding stages of Causal Forcing.
Exposure Bias: Causal CD exhibits less severe exposure bias than causal DMD during autoregressive rollout, and causal CD initialization performs better for subsequent asymmetric DMD.Causal DMD may produce sharper early frames, but later frames drift rapidly and degrade to an unacceptable level.

4. Experiments

Experiments show that frame-wise Causal Forcing++ matches or surpasses prior chunk-wise methods while reducing latency, and that causal CD is the strongest and most efficient initialization across low-step settings.

Performance comparison with existing methods: 84.14 Total and 84.89 Quality make 2-step frame-wise Causal Forcing++ the best VBench-performing method among prior 4-step chunk-wise baselines.The compared baselines are CausVid, Self Forcing, and Causal Forcing.
Performance comparison with existing methods: 50% lower latency and about 1.4× higher throughput are achieved by 2-step frame-wise Causal Forcing++ while preserving comparable overall generation quality.Its 4-step version further improves Quality, VisionReward, and Dynamic Degree over prior SOTA methods while retaining 50% lower latency.
Ablation studies: 5,000 A800 GPU·hours and 1,500 GB of storage make Self Forcing ODE initialization costly, while it underperforms all alternatives across metrics in 2-step and 4-step settings.Its VBench Total remains below 80 and Dynamic Degree reaches at most 2.
Ablation studies: 0 Dynamic Degree, 1.101 VisionReward, and -14 Instruction Following show that direct multi-step AR diffusion initialization nearly collapses in the 1-step setting.Performance improves with more generation steps but remains weak, demonstrating that strong few-step initialization is indispensable.
Ablation studies: Causal CD matches or surpasses causal ODE across all step settings and achieves the best overall performance in the 2-step setting.In 1-step generation, causal CD improves Total, Quality, and Dynamic Degree over causal ODE; in 2-step generation, it has the highest Total, Quality, and VisionReward.
Ablation studies: About 0.5 lower VisionReward for causal DMD than causal CD across settings shows that causal DMD is a weaker few-step initialization replacement.Causal DMD improves over AR diffusion initialization but is usually also worse than causal ODE.

5. Conclusion

Causal Forcing++ replaces causal ODE distillation with causal consistency distillation to improve training efficiency and enable strong frame-wise 2-step performance. It reduces latency by 50% and produces high-quality results compared with alternative initializations.

Conclusion: Causal Forcing++ replaces causal ODE distillation with causal consistency distillation, substantially improving training efficiency.The approach is presented as a response to training inefficiency in existing autoregressive diffusion distillation methods.
Conclusion: 50% latency reduction accompanies performance comparable to or better than prior SOTA methods in the frame-wise 2-step setting.The conclusion identifies this as the first achievement of comparable or better performance under this setting.
Conclusion: Causal CD matches or outperforms causal ODE while dramatically reducing time and storage costs.Self Forcing ODE, multi-step, and causal DMD initializations perform worse in the ablation study.
Conclusion: Causal CD initialization yields high-quality results, whereas causal ODE shows antler separation artifacts and other initializations produce poor visual quality.Visual comparisons after asymmetric DMD report Causal CD as comparable to or better than causal ODE, while multi-step and causal DMD perform worse.