Training on custom dataset conditional

[AnabalonJT]

trafficstars

Hello, I have seen a lot of people having problems with the training on conditional. What I did that worked was to first preprocess the images correctly, they gave as a lot of options like depth, segmentation, etc. You can just pass a picture in rgb and it will be okay the things is to change the configyaml to accept it. Then I couldn't train them together the 2 stages so I had to first train a VQGAN model on it's own, one training for the target image, another training for the conditional, once they can rebuild the image correctly each part and trained tehm together. I made the conditional Vqgan and pass them the ckpt to each stage and then the training went smooth. To the point were ,from just the condtional, it creates a pretty good recreation of it or even create a new one. Just an insight of what I did that worked. I had inspiration on the imagenet edge/ scale preprocess images, But I had concatenated (target/conditional) the images so It want help that much.