A Noise Scheduling Diffusion Model Based on Edge and Semantic Protection

Traditional generative diffusion models typically learn an isotropic Gaussian denoising process, employing an isotropic noise addition and removal strategy when processing each pixel in the image space. Consequently, this approach overlooks the influence of potentially valuable structural information within the image on the denoising and generation processes of the diffusion model. It is well-known that images contain various geometric structures, such as edges and fine details. The isotropic strategy often leads to blurred edges and loss of detail, thereby degrading the overall quality of the generated images. To alleviate this defect, we modulate dynamically the noise via the semantics and edges of the image. Our experiments demonstrate that our method excels at learning mid-frequency and low-frequency information within the dataset, which is crucial for reconstructing the structure and shape of images. Under the Fréchet Inception Distance (FID) metric, the diffusion process of our proposed model outperforms several recently introduced baseline models in the task of unconditional image generation.