nagadomi/waifu2x

waifu2x is an image upscaler and cleanup tool aimed at anime-style art, with a separate model that also handles regular photos. You hand it a small or noisy picture and it returns a larger, cleaner version, using deep convolutional neural networks rather than the simple stretching that built-in image editors perform. The repository points out that active development has moved to a PyTorch successor called nunif, so this codebase is the original Torch-based implementation. It offers three operations: noise reduction at four configurable levels, 2x upscaling, and a combined mode that does both in one pass. You can run it from the command line on a single file, on a list of files for batch conversion, or as a small web application that you open in a browser at a local URL. There is also a worked example in the README for upscaling video, where you extract the frames with avconv/ffmpeg, run waifu2x over each frame, and stitch them back together with the original audio. Switching to the photo model is just an extra command-line flag pointing at a different model directory. You would use it when you have a low-resolution anime image, screenshot, or illustration that you want to enlarge for a wallpaper, print, or asset reuse, and the usual image-editor enlargement looks blocky or smudged. The code is written in Lua on top of Torch7 and expects an NVIDIA GPU with CUDA installed (cuDNN optional for a speed boost). A Docker image is provided, and the README also mentions trying the hosted web demo at waifu2x.udp.jp before installing anything, plus separate Windows-friendly forks like waifu2x-caffe and waifu2x-ncnn-vulkan.