Research map / Image generation / Fast diffusion sampling
$$\Delta $$-DiT: Accelerating Diffusion Transformers without Training via Denoising Property Alignment
Pengtao Chen, Mingzhu Shen, P.D. Ye, Jianjian Cao, Chongjun Tu, Christos-Savvas Bouganis, Yiren Zhao, Tao Chen
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 |
|---|---|---|---|
| cache back blocks instead of front blocks (outline quality) (ablation) | Delta-Cache · PIXART-alpha, 500 MS-COCO2017 | Average gradient: 36.170 -> 59.320 | Improved |
| cache front blocks instead of back blocks (detail quality) (ablation) | Delta-Cache · PIXART-alpha, 500 MS-COCO2017 | PIQE: 19.974 -> 13.137 | Improved |
| apply Delta-Cache to back blocks early and front blocks late (stage-adaptive) | Delta-Cache · MS-COCO2017 | FID: 34.819 -> 35.882 | Mixed |
| stage-adaptive Delta-DiT vs PIXART-alpha at same cost (T=13) | PIXART-alpha · MS-COCO2017 | FID: 39.989 -> 35.882 | Improved |
| Delta-DiT vs PIXART-alpha speedup | PIXART-alpha · MS-COCO2017 | latency (ms): 2290.668 -> 1534.551 | Improved |
| Delta-DiT vs DiT-XL (T=20) | DiT-XL · ImageNet-50k | FID: 15.893 -> 13.289 | Improved |
| Delta-DiT vs PIXART-alpha-LCM | PIXART-alpha-LCM · MS-COCO2017 | FID: 40.433 -> 39.967 | Improved |
| use Delta-DiT with EulerD solver (ablation) | PIXART-alpha · MS-COCO2017 | FID: 39.688 -> 35.735 | Improved |
| use Delta-DiT with DEIS solver (ablation) | PIXART-alpha · MS-COCO2017 | FID: 37.675 -> 35.302 | Improved |
| reverse cache order (front early, back late) (ablation) | Delta-DiT · MS-COCO2017 | FID, IS, CLIP | Hurt |
| increase boundary b from 0 to 20 (ablation) | Delta-DiT · MS-COCO2017 | FID, IS, CLIP | Mixed |
| increase cached blocks Nc from 0 to 28 (ablation) | Delta-DiT · MS-COCO2017 | FID, IS, CLIP | Mixed |
Closest papers by meaning
- Accelerating Diffusion Transformer via Increment-Calibrated Caching with Channel-Aware Singular Value Decomposition (CVPR 2025 · similarity 0.93)
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.92)
- FasterDiT: Towards Faster Diffusion Transformers Training without Architecture Modification (NeurIPS 2024 · similarity 0.92)
- Layer- and Timestep-Adaptive Differentiable Token Compression Ratios for Efficient Diffusion Transformers (CVPR 2025 · similarity 0.92)
- HarmoniCa: Harmonizing Training and Inference for Better Feature Caching in Diffusion Transformer Acceleration (ICML 2025 · similarity 0.92)
- Flash Diffusion: Accelerating Any Conditional Diffusion Model for Few Steps Image Generation (AAAI 2025 · similarity 0.91)
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.
