Last post covered the main technical hurdle to create consistent characters.

I had begun the development of the VN with a known limitation, and I had postponed setting up a good solution, which I knew it existed. The solution is called a Lora (Low Rank Optimization) which will give you a small size plugin model to generate a character, a style, a thing... The results I had seen looked good, but I was a bit worried about how it would affect the style I was looking for.

My graphics card wasn't suited for this type of training, so I would have had to rely in a cloud computing service. How much would it take me to train a character? I didn't know, how many characters was I going to train over the whole development? four or five per act, 15-20, I guessed. What if it doesn't work at first, or the model needs refinement? Wasted money. I was doubtful.

And even then with AI development as fast as it goes, it could be that training requirements could decrease. So I think I made the right decision to, see what I was going to really need.

By the second lesson Louis gives to Claire, I found the real problem with relying solely in the basic model to replicate faces. Once you change setting, the difficulty to replicate a face satisfyingly sky-rockets.

So I had to look into Lora training seriously, and luckily I found out then there was an affordable graphics card with 16GB of VRAM. Twice as much as the 3070 with more or less the same speed. I don't play lots of games anymore, so a sideways upgrade seemed reasonable if I was going to use it for this project.

The upgrade went ok, and I trained my first Lora models on the existing art I had of Claire and Catherine, maybe I threw in even a realistic picture to see what would happen.

It then became apparent, that with the inconsistency of images in the training set, the results wasn't really looking just like the first versions of the characters, but the Lora worked in being consistent with itself which was enough

Only one thing to be sad about: Bye, bye Claire 1.0