K-Culture GlossaryLTechnical words in the news
LoRA (Low-Rank Adaptation)
Low-Rank Adaptation
A low-cost fine-tuning method that trains a small "add-on patch" instead of the whole model — cheap enough to run on a personal PC.
In plain words
Low-Rank Adaptation — a fine-tuning technique that doesn't retrain the whole model. Instead, it trains a small "patch" (an adapter) and attaches it to the original model. It's like remodeling a single room instead of reconstructing the entire building, cutting costs to a tiny fraction of the original.
This is what made it possible to customize models even on a personal PC. It's especially popular in image-generation communities, where people create and share LoRAs for specific art styles or specific characters. That same popularity is why disputes over style imitation and likeness rights tend to swirl around this term.
Fine-tuning vs LoRA: what's the difference
Trains only a small added piece, Retrains the entire model
LoRA is itself a type of fine-tuning, so asking 'which one should I use' is a bit of a false choice. Fine-tuning is the broad category, and LoRA is a cheap way of doing it. When articles today say a model was 'fine-tuned,' it usually actually means LoRA.
| Aspect | LoRA (Low-Rank Adaptation) | Fine-tuning |
|---|---|---|
| What gets trained | Only a small added piece | All the numbers inside the model |
| Cost | Has come down to the point where a personal PC can handle it | Requires enterprise-grade hardware |
| Output | The original stays as is; a single small file is produced | An entirely new full model is produced |
| Swapping | Only the piece is swapped in and out; multiple can be used interchangeably | The whole model has to be replaced |
| Analogy | Adding a patch to existing clothes | Making a new set of clothes |
Rule of thumbFine-tuning is the goal, and LoRA is a cheap way to achieve that goal — it's not a matter of choosing one or the other.
See also
Stories using this term
No story has used this term yet. New ones attach here automatically.