Research map / Image generation / Fast diffusion sampling
DDiT: Dynamic Patch Scheduling for Efficient Diffusion Transformers
Dahye Kim, Deepti Ghadiyaram, Raghudeep Gadde
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 |
|---|---|---|---|
| use third-order instead of first-order finite difference for latent evolution (ablation) | DDiT · COCO | FID: 34.71 -> 33.42 | Improved |
| use third-order instead of second-order finite difference (ablation) | DDiT · COCO | FID: 34.28 -> 33.42 | Improved |
| use third-order instead of first-order finite difference (ablation) | DDiT · DrawBench | CLIP: 0.2927 -> 0.3136 | Improved |
| use third-order instead of second-order finite difference (ablation) | DDiT · DrawBench | ImageReward: 1.0128 -> 1.0284 | Improved |
| raise threshold tau from 0.001 to 0.01 (ablation) | DDiT · DrawBench | ImageReward: 1.0284 -> 1.0124 | Hurt |
| raise threshold tau from 0.001 to 0.01 (ablation) | DDiT · DrawBench | speedup: 2.18 -> 3.52 | Improved |
| lower threshold tau from 0.001 to 0.004 setting (tau=0.004 vs 0.001) (ablation) | DDiT · DrawBench | speedup: 2.18 -> 1.88 | Hurt |
| add DDiT dynamic patching to FLUX-1.Dev (50 steps) | FLUX-1.Dev · COCO | FID: 33.07 -> 33.42 | No clear effect |
| combine DDiT with TeaCache | DDiT · Text-to-image speed | inference time (secs/image): 5.5 -> 3.4 | Improved |
| add TeaCache (delta=0.05) to DDiT (tau=0.001) | DDiT on Wan-2.1 · VBench | VBench score: 80.97 -> 80.53 | Hurt |
| add DDiT (tau=0.001) to Wan-2.1 | Wan-2.1 · VBench | VBench score: 81.24 -> 80.97 | No clear effect |
| use DDiT instead of fewer-step FLUX-1.Dev (28 steps) at similar speed | FLUX-1.Dev · DrawBench | ImageReward: 1.0107 -> 1.0284 | No clear effect |
Later papers that compared with this method
1 later accepted paper reports D-DiT as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Beyond Text-to-Image: Liberating Generation with a Unified Discrete Diffusion Model (ICLR 2026 · better on 2 of 2 shared results)
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
- FlexiDiT: Your Diffusion Transformer Can Easily Generate High-Quality Samples with Less Compute (CVPR 2025 · similarity 0.93)
- Layer- and Timestep-Adaptive Differentiable Token Compression Ratios for Efficient Diffusion Transformers (CVPR 2025 · similarity 0.92)
- Dynamic Diffusion Transformer (ICLR 2025 · similarity 0.92)
- SparseDiT: Token Sparsification for Efficient Diffusion Transformer (NeurIPS 2025 · similarity 0.91)
- D^2iT: Dynamic Diffusion Transformer for Accurate Image Generation (CVPR 2025 · similarity 0.90)
- HarmoniCa: Harmonizing Training and Inference for Better Feature Caching in Diffusion Transformer Acceleration (ICML 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.
