Eliminates the resource-heavy teacher+critic distillation stack (Distribution Matching Distillation) used for few-step causal video diffusion, replacing it with generator-only post-training against a precomputed target distribution. Introduces stochastically truncated clean-exit supervision, a lightweight VAE decoder, and staged vector-Jacobian products to make representation distribution matching memory-feasible for video. Reaches 84.87 on VBench (beating the prior best few-step causal baseline by 0.36) using only 20 generator updates and 16 A100 GPU-hours, cutting both memory use and training time versus the three-network distillation approach.
