Questions about Invertible ISP training

[anArkitek]

Hi, thanks for your great work and your raw coco dataset. I have a question about the details of invertible-isp training.

In the paper you mentioned that you trained the invertible-isp using mit-adobe 5k dataset. To my knowledge, this dataset have each raw image corresponding to five rgb images which are retouched by 5 experts separately. So how did you do the training? Did you only train the rgb->raw direction such that all 5 rgb images are mapped to one raw image, or you just select one rgb randomly and use the raw-rgb pair to train in both rgb->raw and raw->rgb directions?

Thanks!