Diffusion-grounded VideoLLM for Entity-aware temoporal grounding
Diffusion-Grounded VideoLLM is a framework that improves video understanding by integrating structured temporal and entity information before language modeling. It uses a Diffusion Temporal Latent encoder to capture dynamic changes, object-grounded representations to link queries to visual evidence, and a mixed-token input layout for joint reasoning. This leads to superior performance in temporal grounding and video question answering tasks.