Research map / Video / Video generation
RotateAttention : RoPE-Aware Rotation and Range Rectification for INT4 Quantized Attention in Video Generation
Yaofu LIU, Wangli Lan, Jinxi Li, Binhang Yuan, Harry Yang
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 |
|---|---|---|---|
| add range-optimized P quantization (ablation) | INT4 attention without rotation · Wan2.2-T2V | PSNR: 17.829 -> 18.427 | Improved |
| add range-optimized P quantization (ablation) | INT4 attention without rotation · HunyuanVideo-T2V | PSNR: 18.332 -> 18.945 | Improved |
| add range-optimized P quantization (ablation) | INT4 attention without rotation · Wan2.2-I2V | PSNR: 22.828 -> 22.069 | Hurt |
| add half rotation (ablation) | INT4 + Optim-P · Wan2.2-I2V | PSNR: 22.069 -> 22.903 | Improved |
| add half rotation (ablation) | INT4 + Optim-P · HunyuanVideo-T2V | cosine similarity: 0.9712 -> 0.9751 | Improved |
| add half rotation (ablation) | INT4 + Optim-P · Wan2.2-T2V | cosine similarity: 0.95178 -> 0.9492 | Mixed |
| add interleaved rotation (ablation) | INT4 + Optim-P · Wan2.2-T2V | cosine similarity: 0.95178 -> 0.95622 | Improved |
| add interleaved rotation (ablation) | INT4 + Optim-P · Wan2.2-I2V | PSNR: 22.069 -> 22.057 | No clear effect |
| add full Hadamard rotation (ablation) | INT4 + Optim-P · Wan2.2-T2V | PSNR: 18.427 -> 19.211 | Improved |
| add full Hadamard rotation (ablation) | INT4 + Optim-P · HunyuanVideo-T2V | PSNR: 18.945 -> 18.854 | Hurt |
| add learned rotation (orthogonal, 1-2 calibration samples) (ablation) | INT4 + Optim-P · Wan2.2-I2V | PSNR: 22.069 -> 22.713 | No clear effect |
| add half rotation plus Optim-P vs SageAttention-style baseline (ablation) | RotateAttention · Wan2.2-T2V | PSNR: 17.829 -> 18.462 | Improved |
Closest papers by meaning
- SageAttention2: Efficient Attention with Thorough Outlier Smoothing and Per-thread INT4 Quantization (ICML 2025 · similarity 0.87)
- VideoRoPE: What Makes for Good Video Rotary Position Embedding? (ICML 2025 · similarity 0.86)
- ComRoPE: Scalable and Robust Rotary Position Embedding Parameterized by Trainable Commuting Angle Matrices (CVPR 2025 · similarity 0.86)
- Q-VDiT: Towards Accurate Quantization and Distillation of Video-Generation Diffusion Transformers (ICML 2025 · similarity 0.85)
- MonarchRT: Efficient Attention for Real-Time Video Generation (NeurIPS 2026 · similarity 0.85)
- SageAttention: Accurate 8-Bit Attention for Plug-and-play Inference Acceleration (ICLR 2025 · similarity 0.85)
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.
