nagadomi/waifu2x

waifu2x is a tool that takes a small or grainy picture and produces a larger, cleaner version of it. The README focuses on anime-style art but says it also supports photos. Two specific operations are exposed: doubling the resolution of an image, called 2x upscaling, and reducing the visual noise from JPEG compression, at four selectable levels. The two can also be combined in one pass. The approach uses deep convolutional neural networks, which are a kind of machine-learning model that learns image patterns by seeing many examples. The README explains that waifu2x was inspired by an academic paper on image super-resolution using these networks. You can run it as a command-line tool, where one command turns an input PNG into an output PNG, or run it as a small web application that you open in a browser on localhost. There are also instructions for processing every frame of a video by extracting frames, running them through waifu2x, and stitching them back together with the audio. Beyond ready-made models, the README shows how to train your own noise-reduction or upscaling model from a folder of high-quality images. Someone would reach for waifu2x to make low-resolution illustrations or old photos look acceptable when blown up, to clean up artefacts on heavily compressed images, or to upscale animation frames. The codebase is written in Lua and runs on the Torch7 framework using NVIDIA CUDA on an NVIDIA GPU, with optional acceleration via cuDNN. The README also notes that development has since moved to a separate PyTorch-based successor project. The full README is longer than what was provided.