colorization-pytorch

Why we included this project

This is the PyTorch reimplementation of the SIGGRAPH 2017 interactive colorization paper, where a user scribbles a few color hints on a grayscale image and the network fills in the rest from learned priors. It is a training codebase rather than a ready-to-run tool: you prepare the ILSVRC dataset, run the two-stage pass (classification loss first, then regression fine-tuning), and check results with PSNR sweeps across hint densities. Pretrained weights are included, with the original Caffe weights converted to PyTorch, so you can test the model without training from scratch. Expect to integrate it into your own pipeline and pair it with the companion GUI repository for interactive use. For anyone studying hint-conditioned image generation, the architecture and loss design are worth reading closely.