// the find
junyanz/BicycleGAN
Toward Multimodal Image-to-Image Translation
BicycleGAN is a 2017 NeurIPS paper implementation for multimodal image-to-image translation — given one input image, it can generate multiple plausible outputs by sampling from a learned latent space, rather than collapsing to a single deterministic output like pix2pix. It requires paired training data and produces diverse results by combining a cVAE and cLR-GAN objective. This is a research artifact, not a production library.
The core idea is technically sound: the bicycle consistency forces the latent code to actually encode meaningful variation rather than being ignored by the generator. Several pretrained models ship with the repo so you can verify it works before committing to training. The codebase structure is clean and shares a familiar pattern with the pix2pix/CycleGAN repos from the same group, so if you've used those, the option system and training loop will feel familiar. The morphing video generation is a nice touch for visualizing what the latent space actually learned.
Last touched in 2020, targets PyTorch 0.4+, and the README explicitly calls out a legacy branch for 0.1–0.3 — modern PyTorch has moved far past this and there's no indication it runs cleanly on current versions without patching. Requires paired training data, which is a hard constraint that rules out most real-world use cases where you don't have ground-truth correspondences. No pip-installable package, no unit tests, no type annotations — this is a research dump, not something you'd integrate. The evaluation methodology leans on AMT and LPIPS scripts from separate repos, so reproducing the paper metrics requires hunting down specific version flags.