joliGEN

Combines GANs, Diffusion, and Consistency models like CycleGAN, CUT, and Palette, providing flexibility for both unpaired and paired image translation tasks as outlined in the features.

Maintains object classes, masks, and bounding boxes during translation, demonstrated in examples like Mario to Sonic while preserving actions and label boxes in BDD100K datasets.

Includes a Dockerized server with REST API and C++ deployment via DeepDetect, facilitating integration into real-world applications without extensive re-engineering.

Features state-of-the-art models like transformers, HDiT, and hybrid CNN-transformers, enabling high-quality outputs for complex tasks such as object insertion and inpainting.