SSDM: Generated image interaction method based on spatial sparsity for diffusion models

doi:10.1016/j.neucom.2025.129805

SSDM: Generated image interaction method based on spatial sparsity for diffusion models

2025-06-14

0

PRE

AI

DOI：10.1016/j.neucom.2025.129805

Original

Original request for help

Save

Abstract

En 中文

Image interaction methods based on diffusion models are significantly superior to traditional methods. However, due to its slow sampling speed and high computational complexity, it could be more conducive to image editing applications. To address these issues, we propose a generated image interaction method based on spatial sparsity diffusion models (SSDM), which utilizes the spatial sparsity of the differences between the edited and original images to reduce computational complexity and accelerate image generation. It takes sparse block data as a constraint, uses difference masks, converts it into an index, and learns the spatial sparse features of differences to describe the image, thereby reducing the network parameters and computational complexity during the training process. In addition, the overlap-add and overlap-save mechanisms are used to ensure coherence and consistency between different boundaries. It also compensates and trains the sampled feature maps by reusing low-frequency information and introduces Lp-norm to replace Euclidean distance to calculate the loss function, thereby enhancing the reconstruction effect of high-frequency image features. Compared to the original methods, experiments on the LSUN and CelebA-HQ datasets show that the proposed method has achieved better performance of PSNR by improving 1.2 dB, reduced computational complexity by 4.1 x, and increased processing speed by 3.6 x.

Keywords:

Image interaction

Diffusion models

Spatial sparsity

Difference masks

Loss function

SSDM: Generated image interaction method based on spatial sparsity for diffusion models

Abstract

Journal

Organization