The project focuses on creating a custom model for image upscaling to address privacy and cost concerns associated with existing SaaS tools. The developers aim to restore old photos and low-quality images while maintaining their original look, avoiding common issues like excessive sharpening or incorrect handling of film defects.
The team defines the task as supervised image-to-image translation. The model takes low-resolution input images with synthetic noise, such as Gaussian noise or color quantization, which simulate the degradation found in old film or digital sensors. The goal is to produce a high-resolution output that is twice the size of the original.
To achieve high-quality results, the training process utilizes a combined loss function. This includes L1 loss for pixel-level accuracy and perceptual loss based on features from a pre-trained VGG network. This approach ensures that the restored images remain visually consistent with the original source material.
