This was a full deep dive, retagging everything from scratch with modern tagging methods. This means more detailed clothing, labeled perspectives, bangs, and other minor detail tags. This included the proper trick for styles of cropping the best heads in the dataset and training them along with the normal data to increase style accuracy. This was trained in layers that means unfinished and line art was layer 3 (6), spot colors were layer as well as the head shots 2 (8), and completed art was the first layer (10). This was optimized for Illustrious 1.0. Dataset now currently consists of more than 300 data points sitting at a total of 324 images.
Now that I have optimized the four I wanted to my next big goal is making my first trained checkpoint. This will consists of the higher tiers of art from ydbunny, tworship, moralgear, and futon I will probably mix in some at a lesser degree top images from some of my other lora as well, and a "background pack" of images so I'm not losing in background detail. I don't know anything about checkpoints, so I'm just going to treat it like lora training and hopefully It provides a good result (●﹏●).
Along with my normal to-do list I also got requested to-do 2 new styles (by 2 different people), I go overboard for style training I absolutely refuse the stereotypical idea of 20-40 images and I personally try to train all images, easily soaring above 100 most of the time. This combined with my hand tagging preferences (I think it looks better) makes me slow so there some gaps in my normal 1 a week uploads, thank you for your patience.
