Research map / Architectures and efficiency / Attention and transformers
HilbertA: Hilbert-Curve–Aligned Sparse Attention for 2D Structured Data
Shaoyi Zheng, Wenbo Lu, Yuxuan Xia, Shenji Wan
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 central shared region (16x16 vs none) (ablation) | HilbertA · diffusion ablation | LPIPS: 0.592 -> 0.516 | Improved |
| add central shared region (16x16 vs none) (ablation) | HilbertA · diffusion ablation | CLIP-I: 0.871 -> 0.901 | Improved |
| change sliding cycle length from 2 to 4 (ablation) | HilbertA · diffusion ablation | LPIPS: 0.521 vs 0.520 | No clear effect |
| use Hilbert ordering instead of Swin attention at 75% sparsity | HilbertA · Qwen3-VL-8B MMBench | score: 0.645 -> 0.705 | Improved |
| use Hilbert ordering instead of Swin attention | HilbertA · Qwen3-VL-8B OCRBench | score: 606 -> 686 | Improved |
| use 16 tiles instead of 4 tiles | HilbertA · Flux.1-dev 2048x2048 | attention speedup: 2.46x -> 4.16x | Improved |
| use 16 tiles instead of 4 tiles | HilbertA · Flux.1-dev 2048x2048 | FID: 38.4 -> 47.5 | Hurt |
| use HilbertA instead of CLEAR | HilbertA · Flux.1-dev 2048x2048 | end-to-end speedup: 1.37x -> 1.44x | Improved |
| Sparge Attention sparse pattern vs dense | Sparge Attention · Flux.1-dev 1024 | attention latency: 0.48x of dense | Hurt |
| Hilbert token reordering overhead | HilbertA · 1024 and 2048 resolution | overhead vs full attention (%): 7.20 and 1.92 | No clear effect |
| use HilbertA with LoRA on 10k images | HilbertA · FLUX.2-Klein-9B | FID: 26.39 -> 25.67 | Improved |
Closest papers by meaning
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.88)
- Bidirectional Sparse Attention for Faster Video Diffusion Training (NeurIPS 2026 · similarity 0.87)
- ToMA: Token Merge with Attention for Diffusion Models (ICML 2025 · similarity 0.87)
- Parameterized Stripe Attention for Efficient Video Generation (NeurIPS 2026 · similarity 0.87)
- CoreMatching: A Co-adaptive Sparse Inference Framework with Token and Neuron Pruning for Comprehensive Acceleration of Vision-Language Models (ICML 2025 · similarity 0.86)
- Faster Neighborhood Attention: Reducing the O(n^2) Cost of Self Attention at the Threadblock Level (NeurIPS 2024 · 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.
