This post was originally published on this site.

Gently down the Stream…

I’ve been hammering away on the game engine and prepping for a friends and family build in March.

With the new workstation setup to start image model training I did some digging into LoRa and fine tuning methodology. This surfaced a need for a data set I wasn’t accounting for, designed for PPL Training. From Gemini’s deep research:

“The technical solution is the mandatory implementation of Prior Preservation Loss (PPL), a cornerstone of the DreamBooth methodology. PPL acts as a powerful regularization term during fine-tuning. It involves introducing “prior-preservation class images” during training — for instance, hundreds of externally generated images categorized as “fantasy character sheet” or “adventurer portrait.” These generalized images ensure the model retains its understanding of the class of image it is generating while simultaneously learning the specific style imposed by the 250 Aethe images.”

So I’ve begun gathering a base set of fantastic creatures and locations. Turns out text-to-image technologies are not good at fantastic beasts. I’ve tried multiple context formats, but because the training data doesn’t exist for anything that well…doesn’t exist… the models can’t interpolate a ‘horrific 4 armed demon with a face peering malevolently out of its stomach’ into an image. Bummer.

I’ve found an acceptable work around for by utilizing a classic monster approach. This is actually similar to the Witcher series of books which began as tales of fairy tale lore in a gritty realistic setting.

So i’ve been gathering these images for pretraining before I wreck a SD3.5 model with my individualized art style images. This has been fun! I’ve had a lot of stunner images along the way but this one I just had to share. Sometimes those AI just surprise and delight you!