Bernini: Latent Semantic Planning for Video Diffusion

Researchers introduced Bernini, a unified video generation and editing framework.

It uses an MLLM to plan high-level video semantics, then a diffusion renderer to synthesize frames.

The two parts can be trained mostly separately, improving efficiency and modularity.

A segment-aware 3D positional encoding helps the system handle video structure better.

Bernini reports state-of-the-art results on several video generation and editing benchmarks.