Retrained from scratch after pruning 20 of the 35 pics from the original dataset. I didn't upscale the pics this time. I also didn't touch the filewords, which weren't segmented by commas (this was one of my first TIs and I still didn't do that). It doesn't seem to have had a negative impact, so I guess perhaps I'm wrong about how important that part is. In the end, it seems the quality of the images is, by far, the most important thing.

800 steps for this TI. I've checked previous steps and they weren't as accurate, so I'll settle for this one, which in my opinion looks very good. It also gets rid of the saturation that was probably the main issue with the previous version. My settings haven't changed.