Research map / Image generation / Autoregressive and latent generation
Efficient Diffusion Transformer with Step-wise Dynamic Attention Mediators
Yifan Pu, Zhuofan Xia, Jiayi Guo, Dongchen Han, Qixiu Li, Duo Li, Yuhui Yuan, Ji Li, Yizeng Han, Shiji Song, Gao Huang, Xiu Li
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 |
|---|---|---|---|
| replace self-attention with mediator-token attention (n=64) | SiT-S/2 · ImageNet 256x256 no CFG | FID: 58.61 -> 53.57 | Improved |
| use fewer mediator tokens (n=4 instead of 64) (ablation) | SiT-S/2 mediators · ImageNet 256x256 no CFG | FID: 53.57 -> 57.67 | Hurt |
| reduce queries and keys by shrinking W_q and W_k dimensions instead of mediators (ablation) | SiT-S/2 · ImageNet 256x256 no CFG | FID: 58.61 -> 60.02 (r=0.5) | Hurt |
| use mediator-based attention in SiT-XL/2 first four layers with CFG 1.5 | SiT-XL/2 · ImageNet 256x256 CFG 1.5 | FID: 2.06 -> 2.01 | Improved |
| use L1 distance instead of L2 for step-wise mediator switching (ablation) | dynamic mediator schedule · ImageNet 256x256 | FID/sFID vs FLOPs | Improved |
| scale to higher resolution | SiT-B/2 with mediators · ImageNet 512 and 1024 | speed-up: 15.7% (512^2) -> 45.4% (1024^2) | Improved |
| apply mediators at 1024 resolution on SiT-S/2 | SiT-S/2 · ImageNet 1024x1024 | FID: -4.90 FID | Improved |
Closest papers by meaning
- Attention-Driven Training-Free Efficiency Enhancement of Diffusion Models (CVPR 2024 · similarity 0.89)
- Dynamic Diffusion Transformer (ICLR 2025 · similarity 0.88)
- CLEAR: Conv-Like Linearization Revs Pre-Trained Diffusion Transformers Up (NeurIPS 2025 · similarity 0.88)
- DiTFastAttn: Attention Compression for Diffusion Transformer Models (NeurIPS 2024 · similarity 0.88)
- SparseDiT: Token Sparsification for Efficient Diffusion Transformer (NeurIPS 2025 · similarity 0.88)
- Layer- and Timestep-Adaptive Differentiable Token Compression Ratios for Efficient Diffusion Transformers (CVPR 2025 · similarity 0.88)
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.
