CogOmniControl: Reasoning-Driven Controllable Video Generation via Creative Intent Cognition

Researchers introduced CogOmniControl, a reasoning-driven framework for controllable video generation.

It separates intent understanding from video synthesis using a specialized VLM and a controllable video model.

The team also released new benchmarks and evaluation methods to better test creative intent understanding.

Results show stronger performance than existing open-source models on two datasets, especially for sparse or abstract prompts.