How to train with more images in the consistent_character_embedding tutorial

[pdavis68]

Thanks so much for this tutorial. It's fantastic at explaining things.

One thing that's unclear is that you use 25 sample images and I trained my embedding on 25 and it works great in the same model I built the images from (LOFI), but it doesn't work nearly as well in other models. So I'm thinking maybe I need to train with more images. (Maybe not. If not, let me know.)

With the 25 training images I found the 85 or so iterations to be the best. Over 100 started showing exaggerated coloring (saturation?) of the images I produced with it.

I actually have about 100-110 images that I think are pretty good representations and I'm more than willing to create more examples.

Do I just want to do fewer iterations or do I want to change the learn rate? Both? Thanks for any direction you can provide on this.

Also, just a note: In creating my training images, oftentimes, the 512x512 images just didn't look very good, so I did a lot of 1024x1024 images and then scaled them down to 512x512 and they looked much better.