RepForcing: Representation-Space Initialization for Autoregressive Video Distillation
RepForcing improves four-step video generation while removing two stages from the initialization pipeline.
The method trains an autoregressive student by comparing predictions with targets through features from a frozen video diffusion transformer. It reuses existing synthetic videos, replacing intermediate teacher training and causal ODE distillation without generating additional trajectories. The subsequent distribution-matching distillation stays unchanged. After that stage, the authors report VBench Total rising from 84.04 to 84.66 and VisionReward from 6.326 to 10.074 against the reported causal-ODE baseline.
HF Daily Papers' note
The method trains an autoregressive student by comparing predictions with targets through features from a frozen video diffusion transformer. It reuses existing synthetic videos, replacing intermediate teacher training and causal ODE distillation without generating additional trajectories. The subsequent distribution-matching distillation stays unchanged. After that stage, the authors report VBench Total rising from 84.04 to 84.66 and VisionReward from 6.326 to 10.074 against the reported causal-ODE baseline.
HF Daily Papers' note
score 4