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:2605.15141 · 96▲ · cs.CV
View on arXiv →Premium readers get an interactive explainer for this paper — a figure you can poke at, not just read.
What it is
The paper distills bidirectional video diffusion models into few-step autoregressive students for real-time interactive video generation, replacing the expensive causal ODE initialization step of prior work (Causal Forcing) with causal consistency distillation (causal CD). Causal CD gets its training signal from a single online teacher ODE step between adjacent timesteps on real videos, avoiding the need to precompute and store full PF-ODE trajectories.
Why it matters
For anyone building interactive or streaming video models, this cuts the training bill and storage overhead of the initialization stage substantially while moving from chunk-wise 4-step to frame-wise 1-2 step generation, which halves first-frame latency. It makes exploring these aggressive low-latency regimes affordable enough to iterate on without curating and storing large trajectory datasets.
Practical takeaway
If you were using ODE trajectory precomputation to initialize few-step AR video students, you can swap in causal CD to get comparable or better quality with roughly 4x less compute and zero extra storage. Watch also for the finding that score distillation (causal DMD) looks sharper on early frames but drifts badly during rollout, so it is not a good initializer.
Key result
In frame-wise 2-step generation on Wan2.1-1.3B, Causal Forcing++ scores 84.14 VBench Total (vs 84.04 for 4-step chunk-wise Causal Forcing) and 6.661 VisionReward (vs 6.326), while cutting first-frame latency 50% (0.27s vs 0.60s) and Stage 2 training cost about 4x (2,900 vs 11,600 A800 GPU-hours). Gains over the prior SOTA baseline are small in absolute terms; the main win is efficiency and latency, and evaluation uses only 100 prompts for VisionReward and dynamic-degree metrics.
Subscribe
Get the next issue.
Free. One email a week. Unsubscribe any time: no account, no dark patterns.