Скачать или смотреть UltraGen: Native 4K Video Generation Model

In this AI Research Roundup episode, Alex discusses the paper:
'UltraGen: High-Resolution Video Generation with Hierarchical Attention(2510.18775v1)'
This work tackles the scaling limits of diffusion transformer video models, which typically cap at ≤720p due to quadratic attention costs. UltraGen introduces a hierarchical dual-branch attention that splits global and local attention for semantic coherence and high-fidelity detail. It employs spatially compressed global modeling and hierarchical cross-window local attention to cut computation while improving information flow. The result is efficient, end-to-end native 1080p/2K/4K video synthesis, scaling low-res models and outperforming prior approaches.
Paper URL: https://arxiv.org/pdf/2510.18775

#AI #MachineLearning #DeepLearning #VideoGeneration #DiffusionTransformer #HierarchicalAttention #4K #HighResolution