The four main types of LoRAs and what each one is for

Subject LoRAs teach the model to generate a specific person, character, animal, or object. The training data is many photos of the subject in different contexts. The trigger wordTrigger wordA unique, deliberately unusual token used in training captions to name what a LoRA teaches, so that including the word in a prompt activates the learned subject or style. is a unique identifier for the subject. The output is a LoRALoRA (Low-Rank Adaptation)A fine-tuning technique that adapts a base model to a specific style or concept using a small additional file. The standard way to customize open-source models. that can put your subject into arbitrary scenes, at arbitrary angles, with arbitrary lighting, while preserving the subject's identity. This is the most common kind of LoRA and the easiest to train successfully. Most face LoRAs (your face, a celebrity, a fictional character) are subject LoRAs.

The classic failure mode of a subject LoRASubject LoRAA LoRA trained to reproduce one specific person, character, or object across new scenes and poses, the most common and easiest kind to train. is poor identity preservation, the output looks like a generic person rather than the specific subject. This usually means insufficient training data or insufficient training steps. The other classic failure mode is overfittingOverfittingWhen a model memorizes its training data instead of learning general patterns, so it does well on what it has seen and poorly on anything new. A common risk when training LoRAs. to a single pose or expression, the output is your subject but always with the same smile and same head angle. This usually means the training data lacked variation. For commercial work, identity preservation is often the single most important metric, and the best subject LoRAs are trained on 30 or more high-variation images.

Style LoRAs teach the model to apply a specific aesthetic, an art style, a photographic look, a rendering technique, a brand visual identity. The training data is many images that share that aesthetic but have different subject matter. The trigger wordTrigger wordA unique, deliberately unusual token used in training captions to name what a LoRA teaches, so that including the word in a prompt activates the learned subject or style. is a description of the style. The output is a LoRALoRA (Low-Rank Adaptation)A fine-tuning technique that adapts a base model to a specific style or concept using a small additional file. The standard way to customize open-source models. that can render any promptPromptThe text description you provide to a model to specify what you want it to generate. in your style. Examples include: a watercolor LoRA, a Wes Anderson cinematography LoRA, a 1990s editorial fashion photography LoRA, a particular illustrator's brushwork LoRA.

Style LoRAs are harder to train than subject LoRAs because the variation in the training data is by design, you want subject diversity but style consistency, and getting that ratio right is fiddly. Style LoRAs also tend to need higher rankRank (LoRA rank)The inner dimension of a LoRA's two small matrices, which sets how much capacity it has to learn. Higher rank captures more complex behavior but risks overfitting and produces larger files. than subject LoRAs (32 to 64 instead of 16 to 32) because styles are higher-dimensional than subjects. The classic failure mode is style bleeding, the LoRA learns the style and a few of the most common subjects from the training set, and tends to produce those subjects regardless of prompt. The fix is more diverse training data.

Concept LoRAs teach the model a new concept, a thing, an action, a relationship, a category. The training data is many images that show the concept in different forms. The trigger wordTrigger wordA unique, deliberately unusual token used in training captions to name what a LoRA teaches, so that including the word in a prompt activates the learned subject or style. names the concept. Examples: a 'people knitting' LoRALoRA (Low-Rank Adaptation)A fine-tuning technique that adapts a base model to a specific style or concept using a small additional file. The standard way to customize open-source models. that fixes the model's tendency to produce nightmare hands holding malformed yarn, a 'chess board' LoRA that fixes the model's inability to render correct chess piece configurations, a 'specific dance pose' LoRA. Concept LoRAs are often used to patch failure modes of base models, the model knows about the concept but does it badly, and a concept LoRAConcept LoRAA LoRA that teaches or repairs a specific concept, action, or object the base model handles badly, such as correct hands or a particular pose. injects enough corrective examples to fix the problem.

Concept LoRAs are an underused category. They are particularly valuable for commercial and editorial work where you need the model to handle a specific class of subject reliably. If you are running a production workflow that generates a lot of, say, hands holding products, you can train a concept LoRA from a few dozen high-quality examples and dramatically improve the hit rate of your generations.

Action LoRAs teach the model to produce a specific dynamic action or interaction, things diffusion models notoriously struggle with. Examples: a 'two characters hugging' LoRALoRA (Low-Rank Adaptation)A fine-tuning technique that adapts a base model to a specific style or concept using a small additional file. The standard way to customize open-source models., a 'person riding a horse' LoRA (this one is famously hard for base models to get right), a 'specific yoga pose' LoRA. Pose LoRAs are similar but focus on body configuration rather than action. These are usually trained with explicit captions about the action, and they are typically lower rankRank (LoRA rank)The inner dimension of a LoRA's two small matrices, which sets how much capacity it has to learn. Higher rank captures more complex behavior but risks overfitting and produces larger files. (8 to 16) because they are teaching the model a narrow skill, not a broad concept.

Finally, there are effect LoRAs that add a particular visual effect (a film grain LoRALoRA (Low-Rank Adaptation)A fine-tuning technique that adapts a base model to a specific style or concept using a small additional file. The standard way to customize open-source models., a dust particles LoRA, a specific color grade LoRA) and trigger LoRAs that activate a specific behavior when a keyword appears in the promptPromptThe text description you provide to a model to specify what you want it to generate.. These are mostly used as polish steps applied at low strength on top of other generations, rather than as primary creative tools.