The paper introduces a novel method for arbitrary-scale image generation and upsampling using a latent diffusion model and an implicit neural decoder. The proposed method addresses the limitations of existing methods, which often suffer from over-smoothing, lack of diversity, and poor scale consistency. By operating in the latent space, the method achieves efficient and high-quality results at arbitrary scales. The method consists of a pre-trained auto-encoder, a latent diffusion model, and an implicit neural decoder. The latent diffusion process is learned through denoising and alignment losses, with errors backpropagated via a fixed decoder to improve image quality. Extensive experiments on multiple benchmarks show that the proposed method outperforms relevant methods in terms of image quality, diversity, and scale consistency, while also being significantly faster and more memory-efficient. The method is particularly effective in handling the "ill-posed problem" of super-resolution, where multiple high-resolution images can be derived from a single low-resolution image.The paper introduces a novel method for arbitrary-scale image generation and upsampling using a latent diffusion model and an implicit neural decoder. The proposed method addresses the limitations of existing methods, which often suffer from over-smoothing, lack of diversity, and poor scale consistency. By operating in the latent space, the method achieves efficient and high-quality results at arbitrary scales. The method consists of a pre-trained auto-encoder, a latent diffusion model, and an implicit neural decoder. The latent diffusion process is learned through denoising and alignment losses, with errors backpropagated via a fixed decoder to improve image quality. Extensive experiments on multiple benchmarks show that the proposed method outperforms relevant methods in terms of image quality, diversity, and scale consistency, while also being significantly faster and more memory-efficient. The method is particularly effective in handling the "ill-posed problem" of super-resolution, where multiple high-resolution images can be derived from a single low-resolution image.