LoRA (Low-Rank Adaptation)

LoRA works by training a small set of additional weights that are applied on top of a base model like Stable Diffusion XL. Rather than fine-tuning the entire model (which requires significant compute and storage), LoRA modifies only a low-rank subset of the model's attention layers, producing files typically 10-200 MB instead of the base model's 2-7 GB.

For asset management, LoRAs introduce unique challenges. They are part of the generation provenance — the same prompt with different LoRAs produces entirely different results. They need version tracking because LoRA files are updated as creators refine their training data. And they need to be linked to the assets they influenced, creating a dependency graph between model components and outputs that traditional DAMs have no concept of.