Что такое LoRA (Low-Rank Adaptation)?

LoRA was a breakthrough for efficient fine-tuning. Instead of updating all of a model's billions of parameters (expensive, slow, produces a multi-GB file per fine-tune), LoRA freezes the base model and trains tiny low-rank matrices that 'adapt' the model's behavior. The result: fine-tunes that take hours instead of days, produce files of MB instead of GB, can be loaded at runtime, and can be combined (use LoRA-A for character + LoRA-B for style). In Stable Diffusion communities, LoRAs power most custom-character and custom-style work. In LLM fine-tuning, LoRA (and its variant QLoRA) is the default — most enterprise teams use it rather than full fine-tuning.