Video models learn physics from executable code, not guesses: PhysOPSD beats 80K-sample fine-tuning with 3K samples

Video generative models are visually mature but remain brittle on physically plausible dynamics and complex interactions, making targeted post-training the main route to physical capability. Supervised fine-tuning relies on offline data that mismatches the states visited at inference, while reinforcement learning optimizes directly on on-policy rollouts but suffers from sparse, ill-defined physical rewards and credit assignment. On-policy self-distillation casts one model as both student and teacher: the student rolls out on inference inputs and the teacher supervises those rollouts with extra privileged context, so the question becomes how to design effective privileged context.

PhysOPSD chooses executable code as privileged physical context, because an executable program explicitly specifies physical laws, object states and dynamic parameters, forming an editable physical world. Text prompts are inherently declarative, while real reference videos entangle physical laws with appearance, identity and background and offer no parametric control. PhysOPSD-Engine translates real physical scenarios into verifiable simulation programs; execution renders a physical draft and a grounded target prompt, with code generating the underlying dynamics and the rendered draft conditioning the self-teacher.

The abstraction of code becomes dual-level physical privilege: one scene renders a dense draft preserving full geometry and contacts and an abstract draft keeping only object-level motion, trajectories and contact events, forming an intermediate information bottleneck. Distilling directly from a dense draft to a text-only student leaves too wide a privilege gap for internalization, and the teacher may overfit explicit geometry as a visual shortcut. On VideoPhy-2 the abstract-only and dense-only variants each gain at their own level, while the stepped dual-level hierarchy attains the highest gain, indicating that abstract drafts serve as the bridge.

The controllability of code becomes controlled physical variation: paired scenes change only parameters such as friction, restitution, mass, initial velocity or stiffness, while geometry, camera, lighting and initial composition stay fixed. Observing independent physical scenes alone can let the generator overfit canonical motions rather than respond to shifts in physical parameters. Both variants share the same initial Gaussian noise to suppress task-irrelevant rollout variance; on VideoPhy-2 training on independent scenes yields a smaller gain than training on controlled pairs, indicating that explicit counterfactual dynamics internalize physical causality better.

The demonstrated scope is limited to physical dynamics that standard simulation engines can compute stably: rigid-body contacts, elastic and deformable solids, and fluids. Turbulent gas diffusion and thermodynamic phase transitions are excluded for numerical instability and the difficulty of isolating causal parameters. Evaluation covers VideoPhy-2, PhyGround and VBench, the default training budget is 3K samples, and at inference the student runs without teacher or simulation guidance.

The physical scope is explicitly limited to dynamics that simulation can compute, excluding chaotic multi-physics and thermodynamic phase transitions. Gains keep rising as training data grows beyond 3K without a premature plateau, so the default budget is not a saturation point. Evaluation rests mainly on automated metrics and qualitative figures, and judgments of physical plausibility depend on those benchmarks' own criteria.