Anashel

/

rpg

like 294

Hi Judi!

I am trying to erase any reference to know artist and build my own reference structure.

As you can see in my prompt, V2 still has dependencies but V3 is getting much more independent. It's not easy as the current SD 1.5 model had a lot of concept spilling. I didn't manage to get a pipeline working for 2.1. As I understood, a lot of artist were remove from 2.1, which I hope will make it a much better baseline to train my style.

In my case I used Unreal Engine to produce various landscape and style as well as other pipeline like MidJourney to tweak some close-up. Ultimately I hope to rely purely on Unreal Engine as I can generate and automated a vast amount of images. Using Unreal to teach the AI is incredibly powerful for landscape, light and perspective. It is also good for some fabrics, but it get much time consuming when you get face, hair, etc.. That's where I get MidJourney to bridge some gap. The original starting point was a merged pack plus very general training.

With V3 I now have more than 4000 images and 300 000 steps. The training process is broken down in small iterations (ex: leather armor) that hold 50 images with 3k to 5k steps. I train 10 concepts at a time (using the json concept file) and make various control image to see if the model is overtrained. I reproduce this cycle for other components (heavy armor, thief rogue, etc...) or I do a 2nd run on existing concept that did not train enough.

Because of that I tend to keep the training rate low and do more steps. That being said, it is still very rudimentary. I may end-up with a model that can only produce a too limited range. But yesterday we did a stream and try to see if my model could improve the style of other's work; in this case cyberpunk art and pencil art of old western drawing. Surprisingly, it did improve a lot both style, specifically impacted the quality of the clothing details. So when V3 is out, I do invite people to take their current model and merge it (at 0.3 weight) with V3 to see if it's improve their pipeline.

Anyway, feel free to ask more questions!! :)

Here is a sample of RPG V2 vs RPG V3. Don't worry too much for now about the face quality. As I train the clothing, armor, etc... I have to do a polishing on the face aesthetic only at the end.

This is the same prompt, same Steps / CFG and same seed. As you can see, there is a improvement in the clothing details. Now that was for leather. I did also the variations for more thieve likes rogues versus assassin (cloth) style.

Same prompt but this slightly moving your cloth description to a thief / swashbuckler rogue.

Now the biggest impact for V3 is the heavy armor and full plate armor. Same prompt, same steps/cfg, same seed. :)