LoRA (Low-Rank Adaptation)
LoRA, short for Low-Rank Adaptation, is a technique for customizing a large pretrained machine learning model without retraining all of its parameters. Instead of modifying the entire model, LoRA introduces small additional weight matrices that are trained on a specific dataset, then combined with the original model at inference time. This makes it possible to adapt a model to a particular style, subject, or character using comparatively little computing power and storage compared to full fine-tuning.
On AI image and video generation platforms, LoRA files are commonly used to let a base model reproduce a consistent look, such as a specific face, art style, or recurring character, across multiple generations. Because LoRA files are small, they are often shared, sold, or uploaded by users independently of the base model, which raises questions about consent and likeness rights when the training data includes real people. Platforms differ in whether they allow user-uploaded LoRA files, review their contents, or restrict them to vetted, in-house adaptations.
For someone evaluating a platform, the presence and handling of LoRA support indicates both a technical capability and a policy choice. It affects how personalized outputs can be, but also introduces variables around content moderation, since a LoRA trained on unauthorized images can be used to generate material resembling a real individual without that person's involvement or knowledge.
Updated: