Lumos-Nexus: Efficient Frequency Bridging with Homogeneous Latent Space for Video Unified Models

Lumos-Nexus is a two-stage unified video generation framework designed to be more training-efficient.

It first trains a lightweight generator with an understanding block, then uses Unified Progressive Frequency Bridging at inference to transfer generation to a higher-capacity pretrained model in a shared latent space.

The method improves realism and temporal coherence on VBench while preserving reasoning-driven control.

The researchers also introduced VR-Bench, a benchmark for evaluating how well models turn inferred intent into coherent, semantically aligned video.