An Efficient Sparse Blocks Inference Method for Image Editing Based on Diffusion Models

Image editing methods based on diffusion models are significantly superior to traditional methods. However, due to their slow sampling speed, high computational complexity, and weak data generalization ability, these have encountered certain limitations in practical applications. This paper proposes an efficient sparse blocks inference method for diffusion models to address this issue. It also compensates and trains the sampled feature maps by reusing low-frequency information and introduces L
p
-norm to replace Euclidean distance to calculate the loss function, thereby enhancing the reconstruction effect of high-frequency image features. This method takes low-frequency sparse block data inputs as constraints, using the masks of the difference and converting them into indices to achieve the reproduction of high-resolution data. Experiments on the LSUN and CelebAHQ dataset, our method improves the inference speed of DDIM by 2.68 ×, PD by 1.53 × and SDEdit by 5.5 ×, reduces the computational complexity of DDIM by 4.1 ×, PD by 1.7 × and SDEdit by 4.8 ×.