// the find
junyanz/interactive-deep-colorization
Deep learning software for colorizing black and white images with a few clicks.
A SIGGRAPH 2017 research tool that lets you colorize black-and-white photos by dropping color hints on a Qt GUI — you click a spot, pick a color, and the network propagates it plausibly across the image. It's a research demo from Berkeley/CMU, not a production library, and it shows.
The core idea holds up: the local hints network is genuinely good at propagating sparse user strokes into natural-looking colorizations without the muddy averages you get from fully-automatic approaches. The dual-backend design (Caffe for the official SIGGRAPH weights, PyTorch reimplementation for anyone who wants to actually run it today) is pragmatic. The ab color gamut picker in the GUI is a clever UI choice — it constrains user input to perceptually valid colors rather than RGB hex. Adobe licensed the tech for Photoshop Elements 2020, which is about as strong a real-world validation as a research repo gets.
Last commit was 2022, and the Caffe dependency is effectively dead — getting the official model running on a modern system is an archaeology project. The PyTorch backend pins to 0.5.0, which is ancient; nothing here has been tested against anything released in the last four years. No pip package, no ONNX export, no way to use this as a library — it's a standalone Qt app or a notebook, full stop. The training code isn't even in this repo; it lives in a separate colorization-pytorch repo, so if you want to fine-tune on your own data you're stitching two abandoned repos together.