joliGEN
Open-source framework from Toulouse for training your own image-to-image models based on GANs and diffusion.
What it does
joliGEN is a framework for training your own image-to-image models. It supports GANs such as CycleGAN and CUT as well as diffusion and consistency models, with paired or unpaired training data. Typical tasks include inserting or removing objects, turning day shots into night or different weather, and making simulated images look realistic while keeping existing annotations intact. This lets you add rare cases to datasets. Since version 5.0 (October 2026) it also covers video models. It ships with quality metrics such as FID and KID, a server with a REST API and a Docker setup.
Who it suits
ML teams in research and development, for example in industrial vision, automotive or satellite imagery, who want to expand training data for detection models or align simulations with real footage. Without Python and deep learning skills you will not get far.
Data protection for businesses
joliGEN runs on your own hardware or in your own cloud. You decide where images and models are stored; no data goes to the vendor. A data processing agreement is therefore not needed for the software itself, and your images are not used for third-party training. There are no business plans; the vendor Jolibrain supports companies with project work on request.
Limits
Linux is the officially supported system. Training works on a CPU, but for reasonable run times you need an NVIDIA GPU with CUDA. There is no ready-made interface for non-technical users: you prepare datasets and train the models yourself.