Looks and LoRAs
What a LoRA is
A small adapter that teaches a model one specific thing, so you can generate your object or your style rather than something like it.
A LoRA is a small adapter trained on top of a model. It teaches that model one thing specifically: a product, a character, a material, a way of drawing. Once trained it lives in your library, and generating with it means the model produces **your** chair rather than a chair. ## Why a prompt is not enough A prompt describes. An adapter shows. You can write a long paragraph about the exact proportions, the exact finish and the exact way light behaves on your product, and still get something adjacent every time. A dozen photographs of the actual thing settle it, because the model is learning from the object rather than from your description of it. That difference matters most on the two jobs prompts are worst at: - **A specific object** that has to be recognisably itself across twenty images - **A house style** that has to hold across a set nobody will see side by side ## Trigger words A trained LoRA gets a trigger: a short invented word like `ch41r` that means "the thing I trained you on". Put the trigger in a prompt and the adapter engages. Leave it out and you get the base model. That is what lets one prompt reference your trained subject and ordinary words in the same sentence. :::note[Why the trigger is a pseudoword] It has to be a token the model has no existing associations with. Training on the word "chair" would fight everything the model already believes about chairs, where `ch41r` starts empty and holds only what you taught it. ::: ## Two kinds here **Image LoRAs**, trained on images, used by **lora-generate**. Built on FLUX, in two generations that do not mix. **3D style LoRAs**, trained on your own textured models, used by **3D LoRA Generate** and by Trellis 2. These learn a visual language for geometry rather than for pictures, which is unusual enough to be worth knowing exists. ## What they are not A LoRA is not a lock on quality or a guarantee of consistency. It shifts the model strongly toward what it learned and it still produces variation, still misses occasionally, and still needs you to pick. It also inherits the base model's limits. A LoRA on a model that cannot render legible text will not render legible text. :::tip[Worth it above a threshold, not below] Training costs time and credits, and pays back across a set. For three images, prompt carefully. For a campaign, a product line, or anything you will return to for months, train once. ::: Next: [training a LoRA](/docs/using-plnty/looks-and-loras/training-a-lora/).A LoRA is a small adapter trained on top of a model. It teaches that model one thing specifically: a product, a character, a material, a way of drawing.
Once trained it lives in your library, and generating with it means the model produces your chair rather than a chair.
Why a prompt is not enough
Section titled “Why a prompt is not enough”A prompt describes. An adapter shows.
You can write a long paragraph about the exact proportions, the exact finish and the exact way light behaves on your product, and still get something adjacent every time. A dozen photographs of the actual thing settle it, because the model is learning from the object rather than from your description of it.
That difference matters most on the two jobs prompts are worst at:
- A specific object that has to be recognisably itself across twenty images
- A house style that has to hold across a set nobody will see side by side
Trigger words
Section titled “Trigger words”A trained LoRA gets a trigger: a short invented word like ch41r that means “the thing
I trained you on”.
Put the trigger in a prompt and the adapter engages. Leave it out and you get the base model. That is what lets one prompt reference your trained subject and ordinary words in the same sentence.
Two kinds here
Section titled “Two kinds here”Image LoRAs, trained on images, used by lora-generate. Built on FLUX, in two generations that do not mix.
3D style LoRAs, trained on your own textured models, used by 3D LoRA Generate and by Trellis 2. These learn a visual language for geometry rather than for pictures, which is unusual enough to be worth knowing exists.
What they are not
Section titled “What they are not”A LoRA is not a lock on quality or a guarantee of consistency. It shifts the model strongly toward what it learned and it still produces variation, still misses occasionally, and still needs you to pick.
It also inherits the base model’s limits. A LoRA on a model that cannot render legible text will not render legible text.
Next: training a LoRA.