The style of Fonzawa.

The trigger word is fonzawa. The artist's primary OC, Petra, was also tagged and should work relatively well: petra \(fonzawa\), yellow eyes, blue hair, low twintails, ahoge, elf. The dataset is entirely SFW.

I finally built myself a PC with a 5090 (perfect timing, right before the prices skyrocketed), and this LoRA is trained entirely locally. I'll be back to training LoRAs in the nearest future, follow me if you're interested. I've also promised a friend an update to the Katawa Shoujo charpack to include the Nurse -- if you're reading this, don't worry, I remember. I'm also thinking of training Anima updates for my previous LoRAs, that model is very impressive!

As for this LoRA (well, LyCORIS, but I'm gonna stick to the old word for consistency), I've been experimenting with the various training params a bit. I originally assumed I was doing something horribly wrong, but eventually I realized kohya-ss's sd-scripts just generates samples very poorly. So I ended up doing a bunch of runs, out of which I picked two models I enjoyed the most:

• Run 2 epoch 8 seems to be the most faithful to the original style as I interpret it.

• Run 6 epoch 4 seems slightly skewed towards a more neutral style variant, but is more flexible.

Oh, and I've also switched from training on base Illustrious to training on NoobAI V-Pred, something the on-site trainer, as I recall, didn't let me do. If you really want a base Illustrious version of this too, write a comment. I don't know if I want to forward-port my other models to Noob, it seems to handle Illustrious LoRAs well enough, and at this point I'd wager the future belongs to Anima, do I'd rather port to it right away. Still, again, if you really want something updated, write a comment.

If you like the style, please support the original author