Efficient and Training-Free Single-Image Diffusion Models (arxiv.org)

A recent advancement in computer vision introduces an efficient, training-free method for generating images that preserve the internal structure of a single reference image. Unlike traditional approaches that rely on hours of computational training to create diffusion models based on a single image, this new method uses a dataset of image patches at various scales, enabling the creation of an optimal score function for noisy patches through a closed-form denoiser. This innovative approach eliminates the need for extensive neural network training while maintaining high-quality image generation. The significance of this development lies in its ability to achieve state-of-the-art generation quality and diversity without the computational burden typically associated with training single-image diffusion models. The technique supports applications such as unconditional image generation, text-guided stylization, and image retargeting. Notably, it allows for rapid image generation—producing megapixel images in just one second and gigapixel images in a matter of minutes—thus addressing scalability and efficiency, which are critical for the AI/ML community. This method also integrates well with latent space diffusion, paving the way for further innovations in image processing capabilities.