Research map / Image generation / Fast diffusion sampling
DDT: Decoupled Diffusion Transformer
Shuai Wang, Zhi Tian, Weilin Huang, Limin Wang
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 |
|---|---|---|---|
| decouple into condition encoder and velocity decoder | Improved-REPA-B/2 · ImageNet 256x256 (400K steps, w/o CFG) | FID: 19.1 -> 16.32 | Improved |
| decouple into condition encoder and velocity decoder | Improved-REPA-XL/2 · ImageNet 256x256 (400K steps, w/o CFG) | FID: 8.14 -> 6.62 | Improved |
| increase encoder share of layers (8En4De for Base) (ablation) | DDT-B/2 · ImageNet 256x256 (DDT-B/2, 400K steps) | FID: best of 4En8De, 6En6De, 8En4De, 10En2De | Improved |
| use aggressive encoder-decoder ratio 20En4De for Large (ablation) | DDT-L/2 · ImageNet 256x256 (DDT-L/2, 400K steps) | FID: best of 10En14De to 20En4De | Improved |
| replace decoder attention with 3x3 conv (ablation) | DDT-B/2 decoder · ImageNet 256x256 (DDT-B/2, 400K steps) | FID: 16.32 -> 16.96 | No clear effect |
| replace decoder attention with MLP only (ablation) | DDT-B/2 decoder · ImageNet 256x256 (DDT-B/2, 400K steps) | FID: 16.32 -> 24.13 | Hurt |
| add statistic dynamic programming sharing instead of uniform sharing at 3.0x (ablation) | DDT-XL/2 · ImageNet 256x256 with CFG | FID: 1.42 -> 1.40 | Improved |
| share self-condition across steps for 2.6x speedup (uniform) (ablation) | DDT-XL/2 · ImageNet 256x256 with CFG | FID: 1.31 -> 1.36 | No clear effect |
| add improved-baseline techniques (SwiGLU, RoPE, RMSNorm, lognorm) to REPA-B/2 (ablation) | REPA-B/2 · ImageNet 256x256 (400K steps, w/o CFG) | FID: 24.4 -> 19.1 | Improved |
| add improved-baseline techniques to REPA-XL/2 (ablation) | REPA-XL/2 · ImageNet 256x256 (400K steps, w/o CFG) | FID: 7.9 -> 8.14 | No clear effect |
| DDT-XL/2 vs baseline training cost | DDT-XL/2 · Training speed XL/2 | s/it: 0.25s/it -> 0.26s/it | No clear effect |
Closest papers by meaning
- Faster Diffusion: Rethinking the Role of the Encoder for Diffusion Model Inference (NeurIPS 2024 · similarity 0.89)
- Remix-DiT: Mixing Diffusion Transformers for Multi-Expert Denoising (NeurIPS 2024 · similarity 0.89)
- SD-DiT: Unleashing the Power of Self-supervised Discrimination in Diffusion Transformer (CVPR 2024 · similarity 0.89)
- Layer- and Timestep-Adaptive Differentiable Token Compression Ratios for Efficient Diffusion Transformers (CVPR 2025 · similarity 0.88)
- Encoder-Decoder Diffusion Language Models for Efficient Training and Inference (NeurIPS 2025 · similarity 0.87)
- UDT: Reconciling U-Nets and Diffusion Transformers with Data-Adaptive Token Reduction (NeurIPS 2026 · similarity 0.87)
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.
