Research map / Image generation / Fast diffusion sampling
DyDiT++: Diffusion Transformers With Timestep and Spatial Dynamics for Efficient Visual Generation
Wangbo Zhao, Yizeng Han, Jiasheng Tang, Kai Wang, Hao Luo, Yibing Song, Gao Ming Huang, Fan Wang, Yang You
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 Spatial-wise Dynamic Token (SDT) to TDW (ablation) | DyDiT-S lambda=0.5 · ImageNet | FID: 31.89 -> 28.75 | Improved |
| add Timestep-wise Dynamic Width (TDW) to SDT (ablation) | DyDiT-S lambda=0.5 · ImageNet | FID: 70.06 -> 28.75 | Improved |
| add SDT to TDW (ablation) | DyDiT-S lambda=0.5 · Fine-grained avg | FID: 20.93 -> 16.94 | Improved |
| use learnable router instead of random (ablation) | TDW · ImageNet | FID: 124.38 -> 28.75 | Improved |
| use token-level bypass instead of layer skipping (ablation) | SDT · ImageNet | FID: 30.95 -> 28.75 | Improved |
| use token-level bypass instead of layer skipping (ablation) | SDT · Fine-grained avg | FID: 18.29 -> 16.94 | Improved |
| use learned dynamic width instead of manual (ablation) | TDW · Fine-grained avg | FID: 28.10 -> 20.93 | Improved |
| dynamic architecture lambda=0.5 vs static DiT-XL | DiT-XL · ImageNet 256 | FLOPs / FID: 118 -> 57.88 FLOPs, 2.27 -> 2.07 FID | Improved |
| dynamic architecture vs pruning (magnitude) on DiT-S | DiT-S · Fine-grained avg | FID: 35.44 -> 13.98 | Improved |
| dynamic architecture lambda=0.5 vs original DiT-S | DiT-S · Fine-grained avg | FID: 12.27 -> 13.98 | Hurt |
| combine DyDiT with DeepCache interval 5 | DiT-XL · ImageNet 256 | FID: 6.73 -> 3.37 | Improved |
| combine DyDiT with 10-step DPM-solver++ | DiT-XL · ImageNet 256 | FID: 11.66 -> 11.10 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- DiffiT — DiffiT: Diffusion Vision Transformers for Image Generation (ECCV 2024) · better on 0 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
- Dynamic Diffusion Transformer (ICLR 2025 · similarity 0.95)
- D^2iT: Dynamic Diffusion Transformer for Accurate Image Generation (CVPR 2025 · similarity 0.90)
- DiCo: Revitalizing ConvNets for Scalable and Efficient Diffusion Modeling (NeurIPS 2025 · similarity 0.89)
- VDT: General-purpose Video Diffusion Transformers via Mask Modeling (ICLR 2024 · similarity 0.89)
- DC-DiT: Adaptive Compute and Elastic Inference for Visual Generation via Dynamic Chunking (NeurIPS 2026 · similarity 0.89)
- Towards Precise Scaling Laws for Video 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.
