Causal Forcing: Autoregressive Diffusion Distillation Done Right for High-Quality Real-Time Interactive Video Generation

To achieve real-time interactive video generation, current methods distill pretrained bidirectional video diffusion models into few-step autoregressive (AR) models, facing an architectural gap when full attention is replaced by causal attention. However, existing approaches do not bridge this gap theoretically. They initialize the AR student via ODE distillation, which requires frame-level injectivity, where each noisy frame must map to a unique clean frame under the PF-ODE of an AR teacher. Distilling an AR student from a bidirectional teacher violates this condition, preventing recovery of the teacher's flow map and instead inducing a conditional-expectation solution, which degrades performance. To address this issue, we propose Causal Forcing, which uses an autoregressive teacher for ODE initialization to bridge the architectural gap, and then applies the same DMD procedure as in Self Forcing. Empirical results show that our method outperforms all baselines across all metrics, surpassing the SOTA Self Forcing by 19.3\% in Dynamic Degree, 8.7\% in VisionReward, and 16.7\% in Instruction Following. Project page: https://thu-ml.github.io/CausalForcing.github.io/; the code: https://github.com/thu-ml/Causal-Forcing.

Causal Forcing is a method for distilling a multi-step bidirectional video diffusion model into a few-step autoregressive (AR) model. This is a necessary path toward real-time interactive video world models, while the existing method, Self Forcing, suffers from a mathematical flaw. Causal Forcing corrects this issue through three stages: Stage 1 trains an AR diffusion model via teacher forcing; Stage 2 uses the AR diffusion model as the teacher for causal ODE initialization; and Stage 3 applies the same asymmetric DMD as in Self Forcing, with a bidirectional teacher. This substantially improves generation quality.