Research map / Image generation / Fast diffusion sampling
SODA: Sensitivity-Oriented Dynamic Acceleration for Diffusion Transformer
Tong Shao, Yusen Fu, Guoying Sun, Jingde Kong, Zhuotao Tian, Jingyong Su
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 dynamic caching scheduling (DCS) via dynamic programming (ablation) | SODA · DiT-XL/2 DDIM 2.5x | FID: 3.83 -> 2.78 | Improved |
| add unified adaptive strategy (UAS) pruning (ablation) | SODA · DiT-XL/2 DDIM 2.5x | FID: 3.83 -> 2.89 | Improved |
| add UAS on top of DCS (ablation) | SODA · DiT-XL/2 DDIM 2.5x | FID: 2.78 -> 2.75 | Improved |
| add DCS on top of UAS (ablation) | SODA · DiT-XL/2 DDIM 2.5x | FID: 2.89 -> 2.75 | Improved |
| use cosine distance instead of L1 for sensitivity error (ablation) | SODA · DiT-XL/2 DDIM 2.5x | FID: 2.92 -> 2.75 | Improved |
| use cosine distance instead of L2 for sensitivity error (ablation) | SODA · DiT-XL/2 DDIM 2.5x | FID: 2.88 -> 2.75 | Improved |
| increase offline random samples from 100 to 1000 (ablation) | SODA · DiT-XL/2 DDIM | FID: 2.75 -> 2.71 | No clear effect |
| vary pruning hyperparameters lambda and beta (ablation) | SODA · DiT-XL/2 DDIM | FID: Delta FID <= 0.02, Delta IS <= 2.2 | No clear effect |
| use SODA instead of DuCa at similar acceleration | SODA vs DuCa · PixArt-alpha | FID-30k: 28.05 -> 27.33 | Improved |
| use SODA instead of ToCa/DuCa at about 2.5x | SODA vs DuCa · DiT-XL/2 DDIM | FID: 3.05 -> 2.75 | Improved |
| use SODA instead of DuCa at 2.5x | SODA vs DuCa · OpenSora VBench | VBench score: 78.39 -> 78.49 | Improved |
| raise acceleration to 4.00x with DDPM | SODA · DiT-XL/2 DDPM | FID: 2.23 -> 3.33 | Hurt |
Closest papers by meaning
- Accelerating Diffusion Transformer via Increment-Calibrated Caching with Channel-Aware Singular Value Decomposition (CVPR 2025 · similarity 0.90)
- HarmoniCa: Harmonizing Training and Inference for Better Feature Caching in Diffusion Transformer Acceleration (ICML 2025 · similarity 0.89)
- Learning-to-Cache: Accelerating Diffusion Transformer via Layer Caching (NeurIPS 2024 · similarity 0.89)
- FasterDiT: Towards Faster Diffusion Transformers Training without Architecture Modification (NeurIPS 2024 · similarity 0.88)
- BlockDance: Reuse Structurally Similar Spatio-Temporal Features to Accelerate Diffusion Transformers (CVPR 2025 · similarity 0.88)
- Accelerating Diffusion Transformers with Token-wise Feature Caching (ICLR 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.
