Research map / Video / Video generation
Transform Trained Transformer for Accelerating Native 4K Video Generation
Jiangning Zhang, Junwei Zhu, Teng Hu, Yabiao Wang, Donghao Luo, Weijian Cao, Zhenye Gan, Xiaobin Hu, Zhucun Xue, Xiangtai Li, Chengjie Wang, Yong Liu
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 |
|---|---|---|---|
| replace T3 attention with shifted windows (reverse: use T3) (ablation) | T3-Video-T2V-1.3B · 720P, 5K iters | VQA: 67.34 -> 69.37 | Improved |
| use batch size 8 instead of 64 (ablation) | T3-Video-T2V-1.3B · 720P | VQA: 69.37 -> 66.65 | Hurt |
| increase batch size from 64 to 128 or 256 (ablation) | T3-Video-T2V-1.3B · 720P | VQA: 69.37 -> 68.72 | No clear effect |
| use large block ratio (3<ratio<6) layers only (ablation) | T3-Video-T2V-1.3B · 720P | VQA: 69.37 -> 67.14 | Hurt |
| use small block ratio (1<ratio<3) layers only (ablation) | T3-Video-T2V-1.3B · 720P | VQA: 69.37 -> 68.69 | Hurt |
| use lightweight deployment scheme (eVAE, 8 DiT steps) | T3-Video-T2V-1.3B · 720P | VQA: 69.37 -> 67.72 | Hurt |
| transform T3 back to full attention after training (ablation) | T3-Video-T2V-1.3B · 720P | VQA: 69.37 -> 69.51 | No clear effect |
| T3-Video vs UltraGen at 4K | T3-Video-T2V-1.3B · 4K-VBench | VQA: 67.43 -> 71.72 | Improved |
| apply T3 to Wan2.2-T2V-5B | Wan2.2-5B · 4K-VBench | VQA: 47.23 -> 67.40 | Improved |
| apply T3 to Wan2.1-I2V-1.3B | Wan2.1-I2V-1.3B · 4K-VBench | VQA: 59.75 -> 63.6 | Improved |
| LoRA instead of full fine-tuning of T3 | T3-Video-T2V-1.3B · 4K-VBench | VQA: 71.72 -> 70.78 | Hurt |
| apply T3 attention at 4K for DiT | Wan2.1-T2V-1.3B · DiT latency at 2176x3840 | latency: 79,774.4 -> 4,165.8 | Improved |
Closest papers by meaning
- Snap Video: Scaled Spatiotemporal Transformers for Text-to-Video Synthesis (CVPR 2024 · similarity 0.89)
- Training-Free Efficient Video Generation via Dynamic Token Carving (NeurIPS 2025 · similarity 0.87)
- DiTCtrl: Exploring Attention Control in Multi-Modal Diffusion Transformer for Tuning-Free Multi-Prompt Longer Video Generation (CVPR 2025 · similarity 0.87)
- Sparse Video-Gen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity (ICML 2025 · similarity 0.87)
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.87)
- Fast Video Generation with Sliding Tile Attention (ICML 2025 · similarity 0.86)
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.
