Getty 10k

Getty Images 10k Lora

• This LORA was trained on nearly 10k images with highly accurate metadata based captions.

Training Info

So this model used the low frequency/high frequency approach with a increased batch size reduced latent size for generalization.

The Lora was then further refined with Face focused images, a high quality subset of the 10k

I plan on doing a refinement pass on the full dataset as it showed great promise having been trained on 10k unique captions, that likely used a 100k or more unique tokens and billions of combinations