In this paper, we propose a heat diffusion model (HDM) to overcome the limitations of the denoising diffusion probabilistic model (DDPM). While DDPM processes the image as a whole, HDM introduces an attention mechanism between pixels to preserve image details and generate more realistic images, considering that adjacent pixels are more likely to belong to the same object. HDM integrates the discrete form of the two-dimensional heat equation into the diffusion and generative formulas of DDPM to compute the relationship between adjacent pixels during image processing. Experimental results show that HDM generates higher quality samples than models such as DDPM, consistent diffusion model (CDM), latent diffusion model (LDM), and vector quantized generative adversarial network (VQGAN).
