Fixing autoregressive video distillation by enforcing causality

What it does This project is a training recipe for autoregressive video diffusion built on Wan2.1. It inserts a causal initialization stage—either a Causal ODE or Causal Consistency Distillation—before standard asymmetric DMD distillation. The result is a family of few-step generators, including what the authors bill as the first frame-wise one-step and two-step autoregressive video models, that handle both text-to-video and image-to-video in the same frame-wise architecture.

The interesting bit Most distillation methods treat the whole sequence symmetrically, which ignores the obvious: in video, the past causes the future, not the other way around. By baking that directional constraint into the ODE or consistency trajectory, the method reportedly outperforms Self Forcing on visual quality and motion dynamics while keeping the same training budget and inference speed.

Key highlights

• Frame-wise models natively unify T2V and I2V by conditioning on the first latent frame.
• Causal Forcing++ swaps the ODE stage for causal Consistency Distillation, removing the need to curate ODE paired data.
• Checkpoints cover chunk-wise 4-step, frame-wise 4-step, frame-wise 2-step, and frame-wise 1-step, plus a long-video variant.
• Training is three stages: autoregressive diffusion, causal initialization, and asymmetric DMD.
• The inference environment matches Self Forcing, implying drop-in compatibility with existing Wan2.1 setups.

Caveats

• Native generation is hard-capped at 81 frames; the authors explicitly warn that benchmarking the 5-second model on long video is “extremely unfair.”
• The documentation is rich in training scripts but sparse on quantitative metrics or perceptual-study details beyond top-line claims.

Verdict Worth a look if you are distilling Wan2.1 or building real-time interactive video pipelines and need a theoretically grounded initialization. Skip it if you want a turnkey long-video generator without external scaffolding.