alexjc/neural-enhance

Neural Enhance is a Python tool that uses a trained neural network to increase the resolution of photos. You feed it a low-resolution image and it outputs a higher-resolution version, scaled up by 2x or 4x. There are also modes for repairing JPEG compression artifacts or reducing noise without changing the image size. The key thing to understand about how it works: the network is not recovering detail that was actually in the original scene. It is guessing what finer detail probably should be there, based on patterns it learned from a large set of training images. The README describes this as the network hallucinating details. The results look convincing in many cases, but they are a creative interpretation rather than a factual reconstruction of what a higher-resolution camera would have captured. The project ships with pre-trained model files so you do not need to train anything yourself to get started. The recommended way to run it is through Docker, a tool that packages the software and all its dependencies into a self-contained unit. The README provides copy-paste commands to set up a terminal alias called enhance so you can process images without knowing Docker internals. If you have an NVIDIA graphics card, a separate Docker image with GPU support is also available and can reduce processing time from tens of seconds to about five seconds per 1080p image on a capable card. For developers who want to train their own models on a custom image dataset, the README covers the training commands and key parameters in detail. Training requires more setup, including the Lasagne deep learning library and either Theano or a compatible backend, with separate installation guides for Linux, Mac, and Windows. The project is written in Python 3.4 and above. It is a research tool from around 2016 and the README does not mention ongoing maintenance. Pre-trained model files are distributed through the GitHub releases section rather than bundled in the repository itself.