I hope you guys enjoy the model. I've always told people to use adetailer in this. However, I generate locally with stable diffusion forge neo and that adetailer isn't available here or I can't find it. So I will be working on an eye lora to help for series my models are from. If you enjoy the model please leave a positive review!

And down here is where my progress notes for Setanta.

So the feet and the brooch were the hardest part of this. The brooch is just small so that was already an isseue. I'm working on it as I am writing this note so it either failed or succeeded. The feet are a more stressful topic. I've noticed the Fate franchise does this a lot with FGO character design. So, in his FGO art his sandals have mesh covering the upper half of the foot, with the toes exposed. HOWEVER, in his FGO sprite design they removed the mesh to leave just the sandal and the skin exposed. And this in turn influenced all of the fanart out there for Setanta, with some covering the feet with the mesh and some foregoing it altogether, and so on. So what constitutes as the official design? I really don't know and I didn't want to decide. So I kept both in and will let the GPU decide when training it.

I also want to start doing more regarding eyes. I have always generated my art locally using stable diffusion, and the extension they have for adetailer. I couldn't find that one on here or at least with illustrious. I'm going to start trying to improve the base quality of my lora eyes and hopefully make an eye lora someday too.

So today I am at an impasse with Setanta. I have started using batch size 6 for the last view models because I want the model to be more flexible. From what I understand, the higher batch size means the model is able to generalize better. So if I wanted to put a character like Bakugo in a cat suit, it would be able to generate that image easier/better than had I used a lower batch size. The advantage of lower batch size is that is learns the details a lot better, so it produces a more accurate reproduction of the character. The tradeoff there is the ability to generalize easier. The crescent brooch isn't great and its driving me nuts, so this next test I will try batch size 2.

Currently testing epoch 38 of 60 with batch size 2. It's not going so great. Since the brooch is small this reminds me of my Loid lora. I -hated- his lapel pin. I actually had to BUY a file for a 3D model of the lapel pin to create close-up and detailed images of it. It is getting better the more I test it so maybe there is some promise here. And by better I mean, it's an artistic interpretation of the brooch because an accurate reproduction seems out of the question. Also the eye closeups I used in this did not seem to work. It's not horrible. But it's not what I want to consider acceptable either. Maybe creating an eye lora for the future is the way to go. Batch size 2 wasn't bad but it was satisfactory to me either. I'll test again with batch size 6 and adjusted repeats.

So good news with the current test. He's great. The crescent brooch is passable. At least passable enough. HOWEVER, the earrings are burned. Which is fair since they are in every image. So back to editing and then retesting!

It should be better but I adjusted the whole dataset and now the brooch isn't good again. Back to another round of training.

It's not there yet. Training once more with increased repeats.