Research map / Image generation / Fast diffusion sampling
Elastic Diffusion Transformer
Jiangshan Wang, Zeqiang Lai, Jiarui Chen, Jiayi Guo, Hang Guo, Xiu Li, Xiangyu Yue, Chunchao Guo
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 adaptive block skipping (ablation) | E-DiT (Qwen-Image) · Qwen-Image | latency (ms): 2431 -> 1967 (skip only); GenEval 0.895 | Improved |
| add adaptive MLP width reduction on top of block skipping (ablation) | E-DiT (Qwen-Image) · Qwen-Image | latency (ms): 1967 -> 1643; GenEval 0.895 -> 0.857 | Improved |
| add block-wise caching (ablation) | E-DiT (Qwen-Image) · Qwen-Image | latency (ms): 1643 -> 1283; GenEval 0.857 -> 0.853 | Improved |
| use adaptive routing instead of non-adaptive random removal (ablation) | E-DiT (Qwen-Image) · Qwen-Image | GenEval: 0.843 -> 0.853 at similar latency (1514 vs 1283 ms) | Improved |
| directly skip borderline blocks instead of caching (ablation) | E-DiT · Qwen-Image | generation quality | Hurt |
| increase maximum reuse count K of cached features (ablation) | E-DiT block-wise caching · Qwen-Image | generation quality | No clear effect |
| increase cache threshold delta (ablation) | E-DiT block-wise caching · Qwen-Image | generation quality: delta 0.03 gives T=1223ms vs 1319ms at 0.015 | Hurt |
| initialize routers with full capacity instead of random (ablation) | E-DiT (Qwen-Image) · Qwen-Image | GenEval: 0.801 -> 0.853; DPG 78.6 -> 85.4 | Improved |
| apply E-DiT to Qwen-Image (turbo) | Qwen-Image · Text-to-image | latency (ms): 2431 -> 1283; DPG 88.9 -> 85.4 | Improved |
| apply E-DiT to Hunyuan3D-3.0 | Hunyuan3D-3.0 · Image-to-3D | latency (ms): 5012 -> 2587; ULIP-I 0.1446 -> 0.1473 | Improved |
| apply E-DiT to FLUX.1-dev | FLUX.1-dev · Text-to-image | latency (ms): 715 -> 374; GenEval 0.665 -> 0.671 | Improved |
Closest papers by meaning
- Accelerating Diffusion Transformer via Increment-Calibrated Caching with Channel-Aware Singular Value Decomposition (CVPR 2025 · similarity 0.91)
- HarmoniCa: Harmonizing Training and Inference for Better Feature Caching in Diffusion Transformer Acceleration (ICML 2025 · similarity 0.90)
- BlockDance: Reuse Structurally Similar Spatio-Temporal Features to Accelerate Diffusion Transformers (CVPR 2025 · similarity 0.90)
- FasterDiT: Towards Faster Diffusion Transformers Training without Architecture Modification (NeurIPS 2024 · similarity 0.90)
- Layer- and Timestep-Adaptive Differentiable Token Compression Ratios for Efficient Diffusion Transformers (CVPR 2025 · similarity 0.90)
- SparseDiT: Token Sparsification for Efficient Diffusion Transformer (NeurIPS 2025 · 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.
