The course

We split the course into ten short chapters, and the same painter walks with us from start to finish. By the end we'll know what every training parameter does, what to look for in a loss curve, how the recipe shifts between a person, a product and a style, and how to catch an overfit from a single image. The outputs here come from LoRAs we actually trained for this guide; where a real failure would be too subtle to see, we used images clearly labeled as staged.

  1. 1

    What is a LoRA?

    We give a model that already paints beautifully one tiny extra skill, and we can take it off whenever we like. So how do we pull that off?

  2. 2

    So we hand the painter an album

    The painter learns whatever he sees in the album, so we need to think a little about what we fill it with.

  3. 3

    Captions and trigger words

    The notes we leave in the margins of the album decide what the painter calls this new skill and when he reaches for it.

  4. 4

    Steps: how many times should the painter study the album?

    Practice is a fine thing, but too much of it pushes the painter into memorizing (overfitting). Luckily we have the bicycle test!

  5. 5

    Learning rate

    This knob decides how much the painter corrects his brush after each attempt. Too low and the learning never finishes; too high and the painting burns.

  6. 6

    What happens during training?

    What does a single step actually do, how do we read a loss curve, and what are the other knobs on the bench for? We go through them one by one.

  7. 7

    One painter, a drawer full of bracelets

    We trained a cat, a person, a sneaker and a style. So what actually changes in the training when the subject changes?

  8. 8

    Scale, and wearing two bracelets at once

    One slider takes us from invisible all the way to caricature. And every image here is a real output!

  9. 9

    Edit LoRAs: before and after

    Some LoRAs learn a change rather than an object. We hand the painter a sketch and he hands us back a painting.

  10. 10

    So how did we make those pairs?

    We built the sketch-to-painting dataset with an image model and a little light table trick. Let's walk through it step by step.