GeoFlip: Benchmarking Geometry-Sensitive Physics in Video Generation
What it is
GeoFlip is a benchmark for evaluating video generation models on their ability to simulate correct physical outcomes based on initial geometric configurations. It uses 482 paired scenes with controlled geometric interventions to test if models can distinguish between stable and unstable states, such as objects tipping or falling.
The problem it solves
Current video generation models often produce visually realistic but physically incorrect motion, failing to respond appropriately to changes in initial object geometry or placement.
What you could use it for
- Evaluating the physical consistency of image-to-video generation models
- Benchmarking how well AI models respect geometric constraints in physics simulations
- Identifying specific failure modes in video generation like rigid body bending or ignored initial conditions
Why it matters
Provides a rigorous, verifiable standard for measuring the gap between visual realism and physical accuracy in generative AI, guiding the development of models that understand cause-and-effect in 3D space.
Read the source ↗Back to the library →
Summary written by an AI model from the source; check the source before you rely on it.