PhysLoop: Harnessing LLMs' Physics Awareness for Self-Improving Video Generation Models
Organizations: Peking University · ByteDance Seed · Hangzhou Normal University · University of California, Santa Barbara ·
Abstract
Video generation models produce visually realistic videos but still struggle with physical dynamics and causal consistency. Learning from generated videos requires fine-grained physical supervision, yet reliable quality judgments, correction pairs, and controlled variations are difficult to obtain. Large language models (LLMs) possess physical knowledge, but struggle to reliably guide video generation models in producing their training videos. We present PhysLoop, a unified harness designed to support physically grounded self-improvement in video generation models. The harness guides and constrains LLMs throughout video generation, selection, and correction, distilling their physical knowledge into physically consistent videos. The framework supports controllable training-data construction and inference-time video repair. By identifying and correcting failures, PhysLoop marks a key milestone toward recursive self-improvement (RSI) in video generation models with limited human intervention. We release 100K videos spanning diverse, realistic indoor and outdoor scenes, along with generation tools and two evaluation datasets: PhysLoop-Gen and PhysLoop-Repair. We evaluate controllable generation and localized repair on these datasets, respectively, and assess physical behavior on the external Physics-IQ Verified benchmark.
Figures & tables
| Method | Human (Gen) | Model (Gen) | Task | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Phys. | Real. | Causal | Win | PC | SA | IQ | AQ | Gen | Repair | P-IQ | |
| Video Generation Models | |||||||||||
| Gemini Omni | 3.11 | 3.82 | 3.18 | 18.94 | 20.79 | 39.17 | 71.17 | 50.40 | 49.00 | 14.67 | 31.66 |
| H3 | 3.63 | 4.27 | 3.55 | 44.93 | 22.98 | 38.68 | 68.47 | 50.62 | 53.00 | 18.00 | 28.14 |
| Seedance 2.5 | 3.80 | 4.29 | 3.80 | 53.79 | 21.51 | 38.69 | 67.63 | 50.63 | 73.00 | 19.33 | 48.07 |
| Seedance 2.5 | |||||||||||
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
| Category | Representative events | Varied parameters |
|---|---|---|
| Rigid-body motion and contact | Projectile motion, bouncing, sliding, rolling, billiard collisions | Mass, initial velocity, friction, restitution |
| Rigid-body mechanisms | Domino chains, pendulums, Newton’s cradles | Release angle, constraint length, object spacing |
| Elastic and deformable systems | Elastic-body impact, cloth draping, flag motion | Stiffness, damping, Young’s modulus, wind velocity |
| Magnetic interactions | Attraction and repulsion | Field mode, strength, initial separation |
| Fluids | Container transfer, spilling, swirling flow | Initial volume, release height, initial velocity field |
| Fracture | Glass-panel and wooden-crate impact | Impactor mass, launch velocity, impact position, damage threshold |
| Family | Parameter | Values |
|---|---|---|
| Ramp rolling/sliding | Friction (static, dynamic) | |
| Newton’s cradle | Release angle | |
| Elastic-body impact | Young’s modulus | kPa |
| Pinned flag in wind | Wind speed | m/s |
| Fluid volume release | Initial center height | m |
| Glass-panel impact | Impactor speed | m/s |
| Agent | Video model |
|---|---|
| – (text only) | Gemini Omni 1.1 Flash |
| – (text only) | Seedance 2.5 |
| – (text only) | MiniMax H3 |
| Claude Opus 5 | Seedance 2.5 |
| Kimi K3 | Seedance 2.5 |
| GPT-6 Astra | Gemini Omni 1.1 Flash |
Video Generation
Video generation studies models that create and edit moving images from text, images, audio, and structured controls.
See all 383 papers