Continuous-Time Distribution Matching for Few-Step Diffusion Distillation
Tao Liu 1 , Hao Yan 2 , Mengting Chen 2, , Taihang Hu 2 , Zhengrong Yue 2 , Zihao Pan 2 Jinsong Lan 2 , Xiaoyong Zhu 2 , Ming-Ming Cheng 1 , Bo Zheng 2, , Yaxing Wang 3,† 1 VCIP, College of Computer Science, Nankai University 2 Alibaba Group 3 College of Artificial Intelligence, Jilin University https://byliutao.github.io/cdm_page/ https://github.com/byliutao/cdm · 2026
Read Continuous-Time Distribution Matching for Few-Step Diffusion Distillation by Tao Liu 1 , Hao Yan 2 , Mengting Chen 2, , Taihang Hu 2 , Zhengrong Yue 2 , Zihao Pan 2 Jinsong Lan 2 , Xiaoyong Zhu 2 , Ming-Ming Cheng 1 , Bo Zheng 2, , Yaxing Wang 3,† 1 VCIP, College of Computer Science, Nankai University 2 Alibaba Group 3 College of Artificial Intelligence, Jilin University https://byliutao.github.io/cdm_page/ https://github.com/byliutao/cdm (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-06376.txt.
Description
Tao Liu and ten co-authors move distribution matching distillation for diffusion models from fixed timesteps to continuous time. Step distillation speeds image generation, but the leading matching approach supervises only at a few preset points along the sampling path. That sparse anchoring, joined with a mode seeking divergence, tends to leave visual artifacts and over smoothed outputs. GANs or reward models must patch the quality back, the authors argue. Their method removes the need for those patches.
A dynamic schedule of random length enforces matching anywhere along trajectories, not just at anchors. A second objective matches latents extrapolated off the trajectory through the student's own velocity field, preserving fine detail. Across two modern image architectures the method delivered highly competitive fidelity for few step generation. No complex auxiliary objectives were required.
Catalog record
| CREATOR | Tao Liu 1 , Hao Yan 2 , Mengting Chen 2, , Taihang Hu 2 , Zhengrong Yue 2 , Zihao Pan 2 Jinsong Lan 2 , Xiaoyong Zhu 2 , Ming-Ming Cheng 1 , Bo Zheng 2, , Yaxing Wang 3,† 1 VCIP, College of Computer Science, Nankai University 2 Alibaba Group 3 College of Artificial Intelligence, Jilin University https://byliutao.github.io/cdm_page/ https://github.com/byliutao/cdm |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 64 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.06376