TTOM: Test-Time Optimization and Memorization for Compositional Video Generation

Researchers introduce TTOM (Test-Time Optimization and Memorization), a training-free framework that improves compositional video generation in Video Foundation Models during inference. The system uses layout-attention optimization and parametric memory to better align text prompts with generated video outputs, showing strong transferability across different scenarios.

• →TTOM addresses compositional weaknesses in Video Foundation Models without requiring additional training.
• →The framework uses a parametric memory mechanism supporting flexible operations like insert, read, update, and delete.
• →TTOM demonstrates powerful transferability and generalization by disentangling compositional world knowledge.
• →Experimental results on T2V-CompBench and Vbench benchmarks show the framework is effective and scalable.
• →The approach focuses on spatiotemporal layout alignment during inference rather than direct intervention to latents.