Research map / Image generation / Autoregressive and latent generation
Phase-Aligned RoPE for Mixed-Resolution Diffusion Transformer
Haoyu Wu, Jingyi Xu, Qiaomu Miao, Dimitris Samaras, Hieu Le
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 |
|---|---|---|---|
| use phase-aligned attention (PMA) instead of PI-LR | RoPE position interpolation · Wan2.1-1.3B mixed-resolution RoPE interpolation comparison | VBench Total: 71.73 -> 77.04 | Improved |
| use PMA instead of PI-HR | RoPE position interpolation · Wan2.1-1.3B mixed-resolution RoPE interpolation comparison | DOVER Overall: 35.04 -> 75.34 | Improved |
| use PMA instead of YARN | RoPE interpolation · Wan2.1-1.3B mixed-resolution RoPE interpolation comparison | VBench Total: 72.46 -> 77.04 | Improved |
| use PMA instead of NTK | RoPE interpolation · FLUX.1-dev mixed-resolution RoPE interpolation comparison | ImageReward: 0.953 -> 0.978 | Improved |
| add Boundary Expand-and-Replace (n_pad 2/2) (ablation) | PMA · Boundary band ablation (Wan2.1-1.3B) | VBench Total: 74.74 -> 77.04 | Improved |
| add Boundary Expand-and-Replace (n_pad 2/2) (ablation) | PMA · Boundary band ablation (Wan2.1-1.3B) | DOVER Overall: 68.43 -> 75.34 | Improved |
| increase HR boundary padding from 2 to 4 (ablation) | PMA · Boundary band ablation (Wan2.1-1.3B) | VBench Total: 77.04 -> 76.90 | No clear effect |
| use Center Square instead of saliency model for HR regions (ablation) | PMA · Wan2.1-1.3B mixed-resolution RoPE interpolation comparison | VBench Total: 77.04 -> 76.75 | No clear effect |
| use PMA instead of MagCache at 4x | Wan2.1-1.3B acceleration · Wan2.1-1.3B acceleration comparison (about 4x) | VBench Total: 73.88 -> 77.04 | Improved |
| use PMA instead of ToCa at 6x | FLUX.1-dev acceleration · FLUX.1-dev acceleration comparison (6x speedup) | ImageReward: 0.345 -> 0.929 | Improved |
| combine PMA with MagCache | PMA · Wan2.1-1.3B acceleration comparison (about 4x) | Speedup: 4.0x -> 7.8x | Improved |
| combine PMA with DMD 4-step | DMD (4-step) · Wan2.1-1.3B step distillation | Speedup: 15.8x -> 30.7x | Improved |
Closest papers by meaning
- SEGA: Spectral-Energy Guided Attention for Resolution Extrapolation in Diffusion Transformers (NeurIPS 2026 · similarity 0.90)
- LEDiT: Your Length-Extrapolatable Diffusion Transformer without Positional Encoding (NeurIPS 2025 · similarity 0.87)
- CLEAR: Conv-Like Linearization Revs Pre-Trained Diffusion Transformers Up (NeurIPS 2025 · similarity 0.87)
- DiTFastAttn: Attention Compression for Diffusion Transformer Models (NeurIPS 2024 · similarity 0.86)
- Unleashing Diffusion Transformers for Visual Correspondence by Modulating Massive Activations (NeurIPS 2025 · similarity 0.86)
- ExtraVAR: Stage-Aware RoPE Remapping for Resolution Extrapolation in Visual Autoregressive Models (NeurIPS 2026 · 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.
