The base model · 1/4
We just met a wildly talented painter
The painter we met has decades of experience behind him and thousands of paintings to his name.
We're going to ask him for 20 paintings, but when we say 'paint Ilker', what he hands back isn't quite the Ilker we had in mind.
in model terms: base model
The dataset · 2/4
The painter asks us for a few details. (Fair enough)
Talented as he is, he has never met Ilker, so of course he doesn't know him. He asks us for some photos and a little background on who Ilker is.
So we put together an album and hand it over.
in model terms: dataset + captions
Training · 3/4
The painter starts practicing from the album we gave him
The album holds nearly every detail of the person we want painted. The painter studies it, makes attempt after attempt, and tries to learn him as well as he can.
How long he keeps the album, how many times he revisits each photo and so on are decisions we make. Naturally, they shape how well he learns.
in model terms: steps × learning_rate
The LoRA · 4/4
And here comes the magic: the painter works everything he learned into an enchanted bracelet.
From now on, whenever he wears that bracelet, everything he learned from our album comes straight back to his hands.
Before we forget: the painter keeps a drawer full of these bracelets, and he can wear several at once to paint something entirely new!
in model terms: each LoRA is really just a .safetensors file
