LoRA stands for Low-Rank adaptation and the term is often used interchangeably with Custom Models.
In GenAI, fine tuning a pre-trained model to perform within a specific style framework or with a specific product often requires significant computational resources. Traditionally this involves re-training the entire model, which can be slow and costly.
Low-Rank adaptation or LoRAs provide a more efficient solution by allowing you to add a smaller, focused adjustment to the underlying model without requiring retraining. They work by allowing you to create and insert small, trainable elements into the existing model. These adjustments, or 'custom models', fine-tune the model for various style or product-specific uses while keeping the core structure and underlying model intact.
This approach is significantly more efficient - instead of re-training the whole model, the LoRA can focus in on what's needed for a specific task - e.g. generating content in a particular style, while still benefitting from the underlying model's core attributes.
Pencil supports the creation of LoRAs with the AI model Bria.
When do you need a LoRA?
As of mid 2026, LoRA's are no longer commonly used in marketing and advertising domains, although they're still quite popular among the open source AI community. For most image generation, you do not need to train a LoRA. Since the release of the Nano Banana family of image models, frontier base models have been capable of style transfer and style reference, as well as character and object consistency via multiple image references and strong prompt adherence - often moreso than a custom-trained LoRA running on an older model.
The latest image models, like GPT Image 2 and Nano Banana Pro, have taken this trend further. These models can usually interpret a detailed prompt with a small set of reference images attached as style guidance. This has allowed most users to achieve style and subject consistency directly in Image Generation removing the need for the longer process of training a custom style or character LoRA model.
Note: Pencil currently limits each workspace with one style LoRA and one product LoRA.