PhysLoop Uses LLM Guidance to Improve Physics in AI-Generated Video

PhysLoop uses large language models to guide video generation, selection and correction, supporting training-data creation and clip repair at inference time.

Video-generation models can produce convincing scenes while showing objects moving or interacting in physically implausible ways. A clip may look coherent but still depict events without plausible cause and effect. Creating training material with controlled physical behavior can also require close supervision and human review. PhysLoop is designed to coordinate video generation, checking and correction with limited human oversight.

Large language models help plan and guide clip generation, selection and correction.

The framework supports two distinct uses. It can help construct training data with controllable conditions, and it can make localized repairs to generated clips during inference, when a model is producing an output. The framework supports controllable training-data construction and inference-time repair of generated clips.

The project includes a collection of 100,000 videos spanning indoor and outdoor scenes, along with video-generation tools. PhysLoop-Gen evaluates controllable generation tasks, while PhysLoop-Repair evaluates localized video fixes. The researchers also assess physical behavior using the external Physics-IQ Verified benchmark.

The researchers frame PhysLoop as a step toward recursive self-improvement, in which models refine training examples with limited human oversight.