Research map / Image generation / Fast diffusion sampling
DIFFSPARSE: ACCELERATING DIFFUSION TRANSFORMERS WITH LEARNED TOKEN SPARSITY
Haowei Zhu, Ji Liu, Ziqiong Liu, Dong Li, Jun-Hai Yong, Bin Wang, Emad Barsoum
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 two-stage training instead of single-stage (ablation) | DiffSparse · MS-COCO PixArt-alpha | FID-30k: 27.40 -> 26.91 | Improved |
| use attention-based token importance instead of norm (ablation) | DiffSparse · MS-COCO PixArt-alpha | FID-30k: 28.89 -> 26.91 | Improved |
| use attention-based importance instead of similarity (ablation) | DiffSparse · MS-COCO PixArt-alpha | FID-30k: 28.07 -> 26.91 | Improved |
| add DiffSparse allocation on norm-based ranking (ablation) | ToCa-style caching · MS-COCO PixArt-alpha | FID-30k: 29.05 -> 28.89 (-0.16) | Improved |
| add DiffSparse allocation on attention-based ranking (ablation) | ToCa-style caching · MS-COCO PixArt-alpha | FID-30k: 28.35 -> 26.91 (-1.44) | Improved |
| use LPIPS loss instead of L2 (ablation) | DiffSparse · MS-COCO PixArt-alpha | FID-30k: 27.68 -> 26.91 | Improved |
| use LPIPS loss instead of SSIM (ablation) | DiffSparse · MS-COCO PixArt-alpha | FID-30k: 27.46 -> 26.91 | Improved |
| use sparsity interval 0.25 instead of 0.125 (ablation) | DiffSparse · MS-COCO PixArt-alpha | FID-30k: 27.91 -> 26.91 | Improved |
| use sparsity interval 0.25 instead of 1.0 (ablation) | DiffSparse · MS-COCO PixArt-alpha | FID-30k: 28.22 -> 26.91 | Improved |
| add warm-start delta=10 instead of delta=0 (ablation) | DiffSparse · MS-COCO PixArt-alpha | FID-30k: 27.40 -> 26.91 | Improved |
| increase warm-start delta from 10 to 20 (ablation) | DiffSparse · MS-COCO PixArt-alpha | FID-30k: 26.91 -> 26.95 | No clear effect |
| add differentiable learning instead of genetic algorithm search (ablation) | DiffSparse · MS-COCO PixArt-alpha | FID-30k: 27.94 -> 26.91 | 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.
- DiCache — DiCache: Let Diffusion Model Determine Its Own Cache (ICLR 2026) · better on 1 of 1 shared result
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
- SparseDiT: Token Sparsification for Efficient Diffusion Transformer (NeurIPS 2025 · similarity 0.94)
- HarmoniCa: Harmonizing Training and Inference for Better Feature Caching in Diffusion Transformer Acceleration (ICML 2025 · similarity 0.93)
- Layer- and Timestep-Adaptive Differentiable Token Compression Ratios for Efficient Diffusion Transformers (CVPR 2025 · similarity 0.93)
- Learning-to-Cache: Accelerating Diffusion Transformer via Layer Caching (NeurIPS 2024 · similarity 0.93)
- Accelerating Diffusion Transformer via Increment-Calibrated Caching with Channel-Aware Singular Value Decomposition (CVPR 2025 · similarity 0.93)
- TinyFusion: Diffusion Transformers Learned Shallow (CVPR 2025 · similarity 0.92)
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.
