can you imagine this lora be trained on just 3 images? yeah, me neither but it works!

how did this happen?

i generated the initial images with perchance. a front view, side view and rear view with enough coherence to resemble the character. after using comfyui locally to do variations and upscaling i thought "why not try to make a lora?"

i have been thinking about it for a while after training my last lora that "in theory" it should be possible to make a character lora with just 3 images and minimal tagging for trigger words.

this way the lora only has the character and no random clutter that bloats training steps and generations. given that this is only gen 4 of the training and it is already slightly overbaked, suggests that lora training is possible with a tiny dataset and minimal tags for training in a fast and efficient way.

in hindsight, one might also check out the hyper lora paper and the experimental model that can generate lora weights on the fly during generation.