LoRA
A LoRA, short for Low-Rank Adaptation, is a small trainable add-on that customizes a large image model to a specific style, character, product or concept without retraining the whole model.
Fully fine-tuning a modern image model means updating billions of parameters, which is slow and expensive. LoRA takes a shortcut: it freezes the base model and trains only a small set of additional low-rank matrices that nudge the model's behavior. The result is a file that is a tiny fraction of the base model's size but can reliably reproduce a look the base model does not know.
In practice, a LoRA is trained on a modest set of example images, often a few dozen, that share what you want to teach: the same character from different angles, products in your brand style, or a particular illustration technique. At generation time the LoRA is loaded alongside the base model, usually triggered by a keyword and scaled by a strength value.
The strength setting is the main dial. Too low and the effect barely shows; too high and the LoRA overpowers the prompt, reproducing training images instead of following your description. Starting near the middle and adjusting is the usual workflow.
LoRAs are how creators get consistency that prompting alone cannot deliver: the same character across a comic, a uniform brand style across a campaign, or a niche aesthetic on demand. They are most associated with open-weight image model ecosystems, such as the FLUX family available on Arteza.
Frequently asked questions
What is the difference between a LoRA and a fine-tuned model?
A fine-tune modifies the whole model and produces a new full-size model. A LoRA leaves the base model untouched and adds a small adapter on top, so it is far cheaper to train, easy to share, and can be mixed with other LoRAs.
How many images do I need to train a LoRA?
Usable character or style LoRAs are commonly trained on roughly 15 to 50 well-chosen images. Variety in angle, lighting and background matters more than raw quantity.
Can I combine multiple LoRAs?
Often yes, for example a character LoRA with a style LoRA. Each one's strength usually needs lowering when combined, because stacked adapters compete for influence over the output.