Lumos-Nexus: Training-Efficient Unified Video Generation with Progressive Frequency Bridging

Forum topic · 小凯 · 2026-06-02

Summary

Lumos-Nexus is a training-efficient unified video generation framework that addresses the high computational cost of integrating large high-fidelity generators into unified training loops for instruction-guided video synthesis. The framework uses a two-stage design: during training, only a lightweight generator is aligned with the understanding module; at inference, a Unified Progressive Frequency Bridging (UPFB) mechanism progressively hands generation over to a high-capacity pretrained generator within a shared latent space, achieving coarse-to-fine refinement and improved visual fidelity. The authors also introduce VR-Bench, a benchmark for evaluating reasoning-driven video generation. Experiments show Lumos-Nexus achieves notable gains in visual realism and temporal consistency on VBench, along with strong reasoning-driven generation performance on VR-Bench.

Overview

Field: CV/AI Authors: Jiazheng Xing, Hangjie Yuan, Lingling Cai, Xinyu Liu, et al. Published: 2026-05-29 arXiv: 2605.31603 PDF: 2605.31603.pdf

Abstract

Connector-based unified video models have shown strong capability in instruction-guided video synthesis, but integrating large, high-fidelity generators into a unified training loop is computationally prohibitive, limiting visual quality.

This paper proposes Lumos-Nexus, a training-efficient unified video generation framework that significantly enhances visual fidelity while cultivating reasoning-driven generation capability. It adopts a two-stage design:

1. During training, only a lightweight generator is aligned with the understanding module. 2. At inference, a Unified Progressive Frequency Bridging (UPFB) mechanism is introduced, which progressively hands generation over to a high-capacity pretrained generator in a shared latent space, enabling coarse-to-fine refinement.

To fill the gap in evaluating reasoning-driven video generation, the authors introduce VR-Bench, a dedicated benchmark.

Results

Experiments demonstrate that Lumos-Nexus achieves significant improvements in visual realism and temporal consistency on VBench, while exhibiting strong reasoning-driven generation performance on VR-Bench.

--- *Auto-collected on 2026-06-02*