ViRDM proposes a teacher‑ and critic‑free post‑training method for few‑step causal video diffusion that adapts representation distribution matching (RDM) from one‑step image generation to multi‑step video rollouts. The approach tackles three obstacles identified in naïve RDM transfer: a memory‑intractable gradient path, a video‑specific optimization regime, and representation distributions that fail to constrain temporal dynamics. To make RDM feasible, ViRDM couples it with stochastically truncated clean‑exit supervision, employs a lightweight VAE decoder, and uses staged vector‑Jacobian products, which together keep gradient memory tractable during autoregressive generation. It further defines suitable generated‑population and initialization regimes for video RDM and adds a lightweight dynamics regularization term to inject missing temporal constraints. By collapsing the usual three‑network distillation (teacher, critic, generator) into generator‑only updates, ViRDM cuts GPU memory consumption and training time. After only 20 generator updates the model attains a VBench score of 84.87, exceeding the prior best few‑step causal baseline by 0.36 points, while consuming 16 A100 GPU‑hours. Exploratory extensions show the recipe also works with lower causal sampling budgets and for one‑, two‑, and four‑step bidirectional video generation.