Research map / Video / Video generation
SLA: Beyond Sparsity in Diffusion Transformers via Fine-Tunable Sparse–Linear Attention
Jintao Zhang, Haoxu Wang, Kai Jiang, Shuo Yang, Kaiwen Zheng, Haocheng Xi, Ziteng Wang, Hongzhou Zhu, Min Zhao, Ion Stoica, Joseph E. Gonzalez, Jianfei Chen and 1 more
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 |
|---|---|---|---|
| fuse linear attention for marginal blocks into sparse attention (SLA vs Sparse Only) (ablation) | SLA · Ablation (Wan2.1-1.3B) | VA: 64.00 -> 76.96 | Improved |
| add linear attention with sparse attention via learned fusion vs directly summing (L+S) (ablation) | SLA · Ablation (Wan2.1-1.3B) | VA: 29.65 -> 76.96 | Improved |
| use linear attention only (ablation) | Wan2.1 · video gen | VA: 76.78 (full) -> 0.042 | Hurt |
| use softmax instead of elu+1 as linear attention activation (ablation) | SLA · Ablation (Wan2.1-1.3B) | VA: 75.50 -> 76.96 | Improved |
| use softmax instead of hedgehog activation (ablation) | SLA · Ablation (Wan2.1-1.3B) | VA: 74.59 -> 76.96 | Improved |
| reduce k_h from 20% to 5% (ablation) | SLA · Ablation (Wan2.1-1.3B) | VA: 75.81 -> 76.96 with FLOPs 10.65T -> 2.73T | No clear effect |
| raise sparsity to 95% with SLA vs baselines at 85-89% | SLA vs VSA/VMoBa/Sparge · video gen | VA: VSA 55.37, VMoBa 32.33 vs SLA 76.96 | Improved |
| fine-tune with SLA vs full attention | Wan2.1-1.3B · video gen | VA: 76.78 -> 76.96 | No clear effect |
| apply SLA to image generation | LightningDiT · ImageNet 512 | FID: 31.87 -> 31.49 | Improved |
| use SLA kernel vs FlashAttention2 | attention kernel · RTX5090 Wan2.1 | kernel speed: 13.7x forward, 6.8x backward | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- VMoBa — VMoBA: Mixture-of-Block Attention for Video Diffusion Models (ICLR 2026) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.93)
- Bidirectional Sparse Attention for Faster Video Diffusion Training (NeurIPS 2026 · similarity 0.92)
- PISA: Piecewise Sparse Attention Is Wiser for Efficient Diffusion Transformers (NeurIPS 2026 · similarity 0.92)
- DiTFastAttn: Attention Compression for Diffusion Transformer Models (NeurIPS 2024 · similarity 0.91)
- Sparse Video-Gen: Accelerating Video Diffusion Transformers with Spatial-Temporal Sparsity (ICML 2025 · similarity 0.91)
- Parameterized Stripe Attention for Efficient Video Generation (NeurIPS 2026 · similarity 0.90)
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.
