Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Preprint |
| Published: |
2026
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2603.22375 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866908907300978688 |
|---|---|
| author | Cai, Yuren Wang, Guangyi Li, Zongqing Li, Li Liu, Zhihui Su, Songzhi |
| author_facet | Cai, Yuren Wang, Guangyi Li, Zongqing Li, Li Liu, Zhihui Su, Songzhi |
| contents | Diffusion models deliver high-fidelity generation but remain slow at inference time due to many sequential network evaluations. We find that standard timestep conditioning becomes a key bottleneck for few-step sampling. Motivated by layer-dependent denoising dynamics, we propose Multi-layer Time Embedding Optimization (MTEO), which freeze the pretrained diffusion backbone and distill a small set of step-wise, layer-wise time embeddings from reference trajectories. MTEO is plug-and-play with existing ODE solvers, adds no inference-time overhead, and trains only a tiny fraction of parameters. Extensive experiments across diverse datasets and backbones show state-of-the-art performance in the few-step sampling and substantially narrow the gap between distillation-based and lightweight methods. Code will be available. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2603_22375 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Three Creates All: You Only Sample 3 Steps Cai, Yuren Wang, Guangyi Li, Zongqing Li, Li Liu, Zhihui Su, Songzhi Machine Learning Artificial Intelligence Computer Vision and Pattern Recognition Diffusion models deliver high-fidelity generation but remain slow at inference time due to many sequential network evaluations. We find that standard timestep conditioning becomes a key bottleneck for few-step sampling. Motivated by layer-dependent denoising dynamics, we propose Multi-layer Time Embedding Optimization (MTEO), which freeze the pretrained diffusion backbone and distill a small set of step-wise, layer-wise time embeddings from reference trajectories. MTEO is plug-and-play with existing ODE solvers, adds no inference-time overhead, and trains only a tiny fraction of parameters. Extensive experiments across diverse datasets and backbones show state-of-the-art performance in the few-step sampling and substantially narrow the gap between distillation-based and lightweight methods. Code will be available. |
| title | Three Creates All: You Only Sample 3 Steps |
| topic | Machine Learning Artificial Intelligence Computer Vision and Pattern Recognition |
| url | https://arxiv.org/abs/2603.22375 |