Keyboard shortcuts

Press or to navigate between chapters

Press S or / to search in the book

Press ? to show this help

Press Esc to hide this help

LoRA (Low-Rank Adaptation)

Low-Rank Adaptation – a lightweight method for fine-tuning large pretrained models (especially diffusion image models and LLMs) by injecting small trainable low-rank matrices into frozen weights, so you can specialize a model (e.g., to a character) with a tiny fraction of the parameters and compute of a full fine-tune.

LoRA is widely used to train consistent character / style adapters on image-generation models (Stable Diffusion and its forks). A LoRA trained on a character produces a reusable adapter that can be loaded into a base model at inference time to get that character’s look across poses, lighting, and eras while keeping the base model frozen.

On One Small Crawl

LoRA work on the snail is a throughline of the AI generation pipeline (ai-agent-pipeline) – the character-consistency backbone that lets the snail appear across eras while still reading as the same character.