Document including guide with photos on LoRA setup here: https://docs.google.com/document/d/1W0zG4rotGty2S7njO-w4nlrjESZIusXbBUWoH6XsjGI/edit?tab=t.zefal0p89cpn
A LoRA (Low-Rank Adaptation) is a lightweight method for teaching an image model new visual tendencies without retraining the entire system.
Instead of modifying the base model, a LoRA applies a focused bias that alters how the model interprets prompts. When activated, it nudges the model toward a particular way of seeing — changing defaults, expectations, and relationships between elements.
In this workshop, LoRAs are not treated as styles or characters.
They are treated as world rules.
A LoRA does not define a single image.
It defines how images usually behave.
A LoRA does not learn a complete world — it alters the conditions under which images are generated.
It shifts what the model treats as normal, expected, and structurally likely.
A LoRA learns patterns of repetition, not isolated examples.
Visual elements that appear everywhere become assumptions the model builds from.
Elements that appear occasionally become secondary details, not defaults.
Because of this, small and consistent datasets often outperform large, varied ones.
With enough repetition, a LoRA can encode a world condition — changing how scenes are assembled, how elements relate, and what the model expects to exist — rather than merely applying a surface aesthetic.
In this workshop, LoRAs are used to teach models how a world normally operates, not just how it looks.