High-Resolution Image Generation Using Artificial Intelligence and Diffusion Modelling

• Main Authors: Scoles, Amanda, Sionis, Giulia, Otero, Beatriz, Utrera, Gladys
• Format: Conference Proceeding
• Language: English
• Subjects: Artificial intelligence; Computational modeling; Deep Neural Network; Diffusion model; Image synthesis; SR3; Super-resolution; Superresolution; Task analysis; Training
• ISSN: 2377-5750
• DOI: 10.1109/PDP62718.2024.00045

Super-Resolution via Repeated Refinement (SR3) is a state-of-the-art super-resolution algorithm based on diffusion model that can enhance the resolution of images. This method is used to pre-trained models on large datasets and can be used for various tasks without requiring training from scratch. Training SR3 from scratch using the ImageNet dataset involves a complex process that requires substantial computational resources and expertise. The idea is applied the trained SR3 model to new images by feeding the low-resolution inputs and obtaining the high-resolution outputs. It's important to note that training SR3 from scratch is a resource-intensive process that requires powerful GPUs and significant computation time. If you do not have access to such resources, an alternative is to use pre-trained models that are already available and fine-tune them on specific datasets or tasks. The paper shows the result of comparing the resolution of the preprocessed images using a significantly smaller number of images to perform the training with those obtained using the pre-trained model. The results obtained show acceptable results without having to perform on large datasets minimizing the computation time to obtain the resolution of images.