.png){kind=icon}
Return
Partial-Path Rectification in Diffusion Sampling
.png){kind=icon}
PRE
AI
DOI：10.1109/tmm.2026.3718271.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Diffusion models (DMs) have achieved remarkable success in modern generative modeling, but their inherently slow sampling process limits practical deployment. Although recent advances have substantially accelerated sampling, existing methods still constrain the capacity of few-step generation because they typically rely on a pre-trained generator to construct training image pairs that inevitably accumulate errors, thereby degrading modeling performance. To mitigate this issue, we propose a novel partial-path rectification framework that decomposes the full transport path into two sub-paths, enabling the construction of ground-truth image pairs to supervise training. Specifically, the first sub-path is optimized using these data-anchored pairs, while the second sub-path is aligned with the first through a consistency training strategy that enforces velocity-field coherence, yielding a unified and continuous sampling path. With this design, our framework achieves significant one-step generation performance while enabling a smooth trade-off between image quality and sampling efficiency. Extensive experiments on four benchmark datasets demonstrate the effectiveness of the proposed method.
Keywords:
Diffusion models
flow matching
few-step sampling
error accumulation
partial-path rectification
Journal
IF:
9.7
Papers:
5.2K
Citations:
2.8W
Organization
Cited Papers
.png){kind=icon}
ShareSaveShareSave
View more