Here's a bit more BTS on how AI models are made and, believe..
Added 2023-08-01 17:43:07 +0000 UTCHere's a bit more BTS on how AI models are made and, believe me, this only scratches the surface with the amount of work going into training and tuning the AI. Keep in mind this is a new AI platform called SDXL that was just released last week so, while I was getting quite good at the old platform, the community has to start over with the new one and figure out what the best way to 'teach' it is.
Attached are two test sheets and three sample images from the 20 images used to train the JMO AI model, for comparison. The sheets are made kind of like plotting a graph in excel - Each one goes, top to bottom, lowest density of training to highest density of training. Here you can see how the images get really close to matching the model but still manage to fall short in certain areas (i.e. nose not quite right, cheeks not the right shape, etc). Then, as the training gets more dense, it sometimes seemingly loses the plot and starts making "generic attractive mixed woman that kinda looks like JMO" which isn't at all what we want. I'm still exploring options and running more training tests with different settings (as seen in the last attachment) and I'm getting closer with each iteration. Don't even get me started on the tiddies yet, the base checkpoints that SDXL was trained on didn't really include a lot of nsfw stuff and the recent checkpoints that have been released aren't much better in that department, despite doing much much better with face reproduction and consistency. WHEW, I know it's a lot to wrap your head around but it's progress so I'm gonna keep working at it!
