Causal-rCM: A Unified Teacher-Forcing and Self-Forcing Open Recipe for Autoregressive Diffusion Distillation in Streaming Video Generation and Interactive World Models

Research Question and Approach

Autoregressive (AR) video diffusion models are essential for real-time streaming and interactive applications, but they often suffer from exposure bias—a discrepancy between training on ground-truth data and inference on model-generated history. While existing methods like teacher-forcing (TF) and diffusion-forcing (DF) attempt to mitigate this, they often struggle with quality degradation over long sequences. The authors propose Causal-rCM, a framework that treats diffusion distillation as a synergy between forward-divergence (mode-covering) and reverse-divergence (mode-seeking) objectives.

Causal-rCM implements this by using teacher-forcing consistency models (CMs) as an offline, stable initialization strategy. This is followed by self-forcing distribution matching distillation (DMD), which optimizes the model on-policy using its own generated rollouts. To improve efficiency, the authors introduce a custom-mask FlashAttention-2 JVP kernel, enabling the use of continuous-time consistency models (sCM/MeanFlow) that converge 10 times faster than traditional discrete-time approaches.

• Unified Divergence Perspective: The paper demonstrates that teacher-forcing CMs provide the optimal initialization for self-forcing DMD, effectively balancing mode coverage with the high-quality distribution matching required for few-step sampling.
• Algorithmic Efficiency: The implementation of continuous-time CMs via a custom JVP kernel allows for significantly faster training convergence compared to discrete-time counterparts.
• State-of-the-Art Performance: The distilled Wan2.1-1.3B model achieves a VBench-T2V score of 84.63 using only 1 or 2 sampling steps, setting a new benchmark for streaming video generation.
• Interactive Capability: The framework is successfully applied to Cosmos 3, an omnimodal foundation model, enabling action-conditioned generation for physical AI and interactive world models.