A new distillation method keeps fast video-generation models from quietly ignoring physics.
Researchers built DyMD, a framework that adapts Distribution Matching Distillation (DMD) so few-step video models don't sacrifice robot-object motion for visual polish. The team found that DMD's usual re-noising step keeps the teacher model's guidance stuck near motion-deficient rollouts, while rollouts with stronger motion are harder for the critic to fit accurately, a combination that pushes distilled models toward video that looks fine but barely moves the way it should. DyMD counters this with two additions: a re-noising schedule that adapts to each rollout's current motion fidelity, and a critic-training method that upweights the harder, higher-motion rollouts instead of averaging them away. Applied to a 14 billion parameter teacher model, DyMD produced a four-step 1.3 billion parameter student with no extra components needed at inference.
On embodied-video benchmarks, the distilled student improved task adherence by 9.6 percentage points on R-Bench and domain score by 5.1 points on PAI-Bench-G versus a standard DMD baseline, while matching its visual quality. More strikingly, when used as a backbone for downstream robot action planning, it more than doubled task success on WorldArena benchmarks: 34 percent versus 16 percent for base DMD. That gap matters because embodied AI systems don't just need video that looks plausible, they need video whose motion is accurate enough for a robot to plan around.
This is a narrow, technical fix, but it's a useful reminder that compressing generative models for speed tends to quietly discard exactly the details, motion, dynamics, cause and effect, that make them useful for anything beyond a demo reel.