arrow
返回

Training-Free Diffusion-Based Conditional Generation with Collaborative Samples

delete2026-10-23
delete0
PRE
AI
S
Siyu Xing
J
Jie Cao
H
Haichao Shi
X
Xiaoyu Zhang
赫然 封面图
赫然 (Ran He)
DOI:10.1109/tip.2026.3739581delete
delete原文链接
delete分享
delete收藏
摘要

摘要

En 中文
近期无条件扩散模型在条件生成任务中展现出卓越性能。通常,扩散模型作为先验,通过引入可微分损失函数参与任务而无需重新训练。传统上,采样过程中输出相互独立,这限制了解空间的探索并忽略了条件分布在样本层面的潜在结构。为应对此挑战,我们提出S-diff,一种基于Stein变分梯度下降(SVGD)的新型即插即用框架。该框架通过在扩散采样过程中探索样本间的相互作用来提升条件生成任务性能。具体而言,我们将多个样本视为SVGD更新中的“粒子”,结合扩散模型在第一时间步的得分函数与现成的可微分损失函数来近似条件得分函数。随后,该条件得分通过其他粒子定义的核函数进行平滑,共同引导粒子向更高概率区域移动。值得注意的是,由于扩散模型的内在属性,近似误差保持良好界限。该框架可无缝集成至常见的连续扩散采样过程和分离采样过程中。我们的实验证明了S-diff在广泛条件生成任务中的有效性。例如,我们在一个具有挑战性的基准上对16个条件生成任务全面评估了S-diff,涵盖视觉、音频和分子领域。结果显示,S-diff平均相对提升9.4%。
Keyword:
Generative model
diffusion model
conditional generation
Stein variational gradient descent

期刊

IEEE Transactions on Image Processing 封面图
IEEE Transactions on Image Processing
IF:
13.7
论文数:
1.1W
被引数:
10.8W

机构

I
Institute of Automation Chinese Academy of Sciences
学者数:
162
论文数: 69
被引数: 3
引用论文

引用论文

TFG: Unified Training-Free Guidance for Diffusion Models
err2024-01-01
err0
PREAI
errYe,Haotian; Lin,Haowei; Han,Jiaqi; Xu,Minkai; Liu,Sheng; Liang,Yitao; Ma,Jianzhu; Zou,James; Ermon,Stefano
err分享
err收藏
Applying Guidance in a Limited Interval Improves Sample and Distribution Quality in Diffusion Models
err2024-01-01
err0
PREAI
errKynkäänniemi,Tuomas; Aittala,Miika; Karras,Tero; Laine,Samuli; Aila,Timo; Lehtinen,Jaakko
err分享
err收藏
Universal Guidance for Diffusion Models
err2023-06-01
err0
errOAAI
errArpit Bansal; Hong-Min Chu; Avi Schwarzschild; Soumyadip Sengupta; Micah Goldblum; Jonas Geiping; Tom Goldstein
err分享
err收藏
Dynamical regimes of diffusion models
err2024-11-17
err0
errOAAI
errBiroli, Giulio; Bonnaire, Tony; de Bortoli, Valentin; Mezard, Marc
err分享
err收藏
Quantum chemistry structures and properties of 134 kilo molecules134千分子的量子化学结构和性质
err2014-08-05
err0
errOAAI
errRaghunathan Ramakrishnan; Pavlo O. Dral; Matthias Rupp; O. Anatole von Lilienfeld
err分享
err收藏
err分享
err收藏
Solving Linear Inverse Problems Provably via Posterior Sampling with Latent Diffusion Models
err2023-01-01
err0
PREAI
errRout,Litu; Raoof,Negin; Daras,Giannis; Caramanis,Constantine; Dimakis,Alex; Shakkottai,Sanjay
err分享
err收藏
ImageNet Large Scale Visual Recognition ChallengeImageNet大规模视觉识别挑战
err2015-04-11
err2.7W
PREAI
errRussakovsky, Olga; Deng, Jia; Su, Hao; Krause, Jonathan; Satheesh, Sanjeev; Ma, Sean; Huang, Zhiheng; Karpathy, Andrej; Khosla, Aditya; Bernstein, Michael; Berg, Alexander C.; Fei-Fei, Li
err分享
err收藏
Improving Diffusion Inverse Problem Solving with Decoupled Noise Annealing
err2025-06-10
err0
PREAI
errZhang,Bingliang; Chu,Wenda; Berner,Julius; Meng,Chenlin; Anandkumar,Anima; Song,Yang
err分享
err收藏
学者 查看更多内容