Research map / Video / Video generation
UltraGen: High-Resolution Video Generation with Hierarchical Attention
Teng Hu, Jiangning Zhang, Zihan Su, Ran Yi
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| native high-res generation with UltraGen instead of Wan direct generation | Wan-1.3B · 1080P text-to-video | HD-FVD: 821.54 -> 214.12 | Improved |
| native UltraGen vs Wan+SR two-stage pipeline | Wan-1.3B · 4K text-to-video | HD-FVD: 471.56 -> 424.61 | Improved |
| native UltraGen vs Wan+SR on prompt following | Wan-1.3B · 1080P text-to-video | CLIP-L: 0.2747 (Wan+SR) vs 0.2654 (UltraGen) | Hurt |
| UltraGen vs non-SR baselines on CLIP | Wan-1.3B · 1080P text-to-video | CLIP-L: 0.2528 (Wan) -> 0.2654; 0.2636 (HunyuanVideo) | Mixed |
| UltraGen efficient attention vs Wan full attention for inference time | Wan-1.3B · 4K generation | inference time: 8h 46min -> 1h 50min (x4.78) | Improved |
| UltraGen vs Wan for 1080P inference time | Wan-1.3B · 1080P generation | inference time: 35 min -> 13 min (x2.69) | Improved |
| remove global attention branch (ablation) | UltraGen · qualitative ablation (goldfish prompt) | visual quality | Hurt |
| remove cross-window local attention or hierarchical attention (ablation) | UltraGen · qualitative ablation | boundary consistency | Hurt |
| remove domain-aware LoRA (ablation) | UltraGen · qualitative ablation | visual quality | Hurt |
| replace hierarchical cross-window attention with Swin attention (ablation) | UltraGen · qualitative ablation | semantic coherence | Hurt |
Closest papers by meaning
- ScaleCrafter: Tuning-free Higher-Resolution Visual Generation with Diffusion Models (ICLR 2024 · similarity 0.81)
- GenTron: Diffusion Transformers for Image and Video Generation (CVPR 2024 · similarity 0.80)
- Ultra Flash: Scaling Real-Time Streaming Video Generation to High Resolutions (NeurIPS 2026 · similarity 0.79)
- IllumiCraft: Unified Geometry and Illumination Diffusion for Controllable Video Generation (NeurIPS 2025 · similarity 0.79)
- Grid Diffusion Models for Text-to-Video Generation (CVPR 2024 · similarity 0.78)
- Re-ttention: Ultra Sparse Visual Generation via Attention Statistical Reshape (NeurIPS 2025 · similarity 0.78)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
