Oleksandr Kuzmenko, EditorPublished 6 min readAI-assisted · editor-reviewed
TL;DR
- Adding structured physics_reasoning and physics_negative_prompt fields lifts video model fidelity without retraining.
- Cosmos3-Nano equipped with Physis-Lang outperforms Google Veo 3.1 on the VideoPhy-2 Hard physics benchmark (62.36 vs 58.43).
- Distilling prompt upsampling into a 4B VLM (PhysThinker) reduced prompt preparation costs from $24.12K to $0.12K.
Key facts
- Physics-IQ Verified Score (Cosmos3-Super)
- 48.2 ± 1.4 (Rank 1)
- VideoPhy-2 Hard Split
- 62.36 (Cosmos3-Nano) vs 58.43 (Veo 3.1)
- Prompting-Only Gain (PhyGenBench)
- +5.62 points on frozen Cosmos3-Nano
- Prompt Pipeline Cost Reduction
- From $24.12K (API) to $0.12K (PhysThinker)
Why Video World Models Defy Physical Laws
Most video diffusion architectures memorize visual co-occurrences without understanding causal constraints, leading to artifacts like solids melting without heat or water flowing uphill. Rather than introducing complex auxiliary latents, Physis-Lang uses structured natural language as the common representation across data curation, fine-tuning, and prompt conditioning.
The Structured Prompt Schema
The framework expands conventional prompts into two explicit conditioning components:
1. physics_reasoning: Explicitly spells out entities, causal mechanisms, interactions, physical principles (e.g., gravity, thermodynamics), and temporal transitions. 2. physics_negative_prompt: Formulates scene-specific implausible failure cases to steer the model away from degenerate physical states.
{
"prompt": "A ceramic coffee mug drops onto a solid hardwood floor",
"physics_reasoning": "Gravitational acceleration pulls the mug downward; impact force exceeds tensile strength of glazed ceramic; brittle fracture propagates from point of contact; shards scatter outward with kinetic energy.",
"physics_negative_prompt": "Mug bouncing like rubber; mug passing through hardwood floor; mug shattering before contact; zero-gravity floating shards"
}Benchmark Results and Frozen Model Gains
On the Physics-IQ Verified snapshot, Cosmos3-Super with Physis-Lang achieved 48.2 ± 1.4, taking the top spot. On VideoPhy-2 Hard split, Cosmos3-Nano scored 62.36 against Google Veo 3.1's 58.43.
Crucially for prompt engineers, running structured physics reasoning and negative prompts on a frozen Cosmos3-Nano without any fine-tuning lifted PhyGenBench scores by +5.62 points (61.67 to 67.29).
Try it in 2 minutes
{
"prompt": "A block of ice melting on a hot steel griddle",
"physics_reasoning": "Thermal conduction transfers heat from the steel surface to the ice bottom; solid water undergoes phase change into liquid; gravity pulls melted water outward across the surface.",
"physics_negative_prompt": "Ice maintaining cube edges while shrinking; water floating upward; sudden vanishing without puddle formation"
}When to use
- Generating videos with complex physical dynamics such as fluids, collisions, or thermal phase changes.
- Formulating negative prompts to prevent visual clipping and defying gravitational laws in AI video.
- Curating synthetic video training sets where physical realism is required.
What to do today
- Incorporate a physics_reasoning block in generative video prompts explaining physical transitions step by step.
- Add an explicit physics_negative_prompt listing physically impossible behaviors like clipping or gravity defiance.
- Test dual-prompt conditioning on local video generators like Wan2.1 or Cosmos3.
Step names are taken from What to do today.
Step names are taken from What to do today. Incorporate a physics_reasoning block in generative video prompts explaining physical transitions step by step.. Add an explicit physics_negative_prompt listing physically impossible behaviors like clipping or gravity defiance.. Test dual-prompt conditioning on local video generators like Wan2.1 or Cosmos3.
- Incorporate a physics_reasoning block in generative video prompts explaining physical transitions step by step.
- Add an explicit physics_negative_prompt listing physically impossible behaviors like clipping or gravity defiance.
- Test dual-prompt conditioning on local video generators like Wan2.1 or Cosmos3.
