AI models for image creation

LoRA: specialising a model to hold a style or a character

A LoRA adapts a model's behaviour without retraining it: a style, a character, a material, a visual identity. What it is, what it depends on, when it is useful (the continuity of a series, of a brand) and what it does not replace.

A LoRA is a method that adapts the behaviour of an image model without retraining all of its parameters: a small file, trained on a chosen set of images, that teaches the model a style, a character, a material or a visual identity. It is not a standalone engine: it depends on the base model it was trained for. For a creative, it serves above all continuity, finding the same aesthetic or the same character across a whole series without starting from scratch; it holds an identity, it does not replace art direction.

What is a LoRA?

A LoRA (Low-Rank Adaptation) is a method that adapts a model's behaviour without retraining all of its parameters. The result is a lightweight file that is added to the base model and steers what it produces.

In image creation, a LoRA serves to introduce or reinforce a characteristic: a style, a character, a material, an aesthetic, a garment, a visual family, sometimes a particular behaviour of the model.

This page says what a LoRA is and when it is useful. How to build one that holds up in production (the image set, the captions, the tests) is in LoRA workflow.

A LoRA depends on the base model

A LoRA is not a standalone engine. It is trained for a specific architecture or family and works only with it: a LoRA for one generation of Stable Diffusion does not apply to FLUX, and vice versa.

Before using one, you therefore need to know its base model, the training method, the original image set, the trigger word if there is one, and the strength at which to apply it. A LoRA downloaded without this information is a file you cannot predict the behaviour of.

When a LoRA is useful

For a creative, the LoRA becomes interesting when it serves to build visual continuity:

  • a series: the same rendering, the same light, the same aesthetic across every image of a collection;
  • a brand: a visual identity that holds from one campaign to the next, without redescribing it every time;
  • a character or a product: the same face, the same object, recognisable from one image to the next and from one scene to the next.

That is what text alone cannot make last, and what a visual reference holds over a few images but not over months of production.

What a LoRA does not replace

A LoRA holds an identity. It does not replace art direction: it does not say which image to make, nor whether it is good. Nor does it replace the method: the choice of engine, the references, retouching by area, the frozen layer. And a LoRA trained on a bad set of images faithfully reproduces that bad set.

The other way to hold an identity, more flexible and without training, is the visual reference you give a role to: Workflow with visual references. The two combine: the reference for a project, the LoRA for what has to hold over time.

The LoRA in the workflow

Once built, the LoRA becomes a building block of the workflow, not the workflow: it is added to the engine, combines with the references, and what it produces is judged, corrected and frozen like everything else. What a LoRA asks of the machine and where it is stored: How much VRAM to generate AI images? and Models in ComfyUI.

At Axiris, building a LoRA is part of setting up a workflow for a studio or a brand that needs visual continuity. The other families of engines: AI models for image creation.

Contact request: a first conversation, free and with no commitment. Every practice is different; it is by listening to you that we propose what suits you.