Distilling a large pretrained video diffusion model into a compact autoregressive generator usually requires a two-stage recipe: initialize the student with an intermediate teacher, then apply distribution matching distillation (DMD). The intermediate teacher supplies causally compatible targets, but the authors found that causal ODE initialization still leaves fine details unresolved, even though it reduces blur.
RepForcing changes the initialization stage. Instead of regressing the student against intermediate-teacher trajectories, it compares clean predictions and training targets in the normalized feature space of a frozen video diffusion transformer. This representation-space supervision reuses synthetic videos already generated by a bidirectional model, so no additional causal ODE trajectories are needed. The subsequent asymmetric DMD stage is unchanged.
Experiments with four sampling steps show that RepForcing preserves generation quality better than intermediate-teacher-based initialization before DMD. After DMD, the four-step chunk-wise model improves VBench Total from 84.04 to 84.66 and VisionReward from 6.326 to 10.074, a 59.25% gain over the reported causal-ODE baseline.