Janus 1.3B uses differerent visual encoders for understanding and generation.
Here, by feeding the model an image and then asking it to generate that same image, we visualize the model's ability to translate input (understanding) embedding space to generative embedding space.