UNIT

Why we included this project

UNIT learns to map images between two visual domains, day scenes to night or snowy landscapes to summer, without needing paired examples. Its shared latent space assumption is the core idea, and the code still works as a clean reference for it, which is why the NIPS 2017 paper keeps getting cited. Be aware it was written for Python 2.7 and the authors have moved on, pointing users to the faster Imaginaire rewrite and to MUNIT for many-to-many translation. The license is also non-commercial, so treat it as a study tool rather than something to drop into a product. For that, the bundled tutorials and pre-trained Synthia-to-Cityscapes model still make a good on-ramp to how classic unsupervised translation pipelines work.