How to train a character LoRA for your AI influencer
Updated · 6 min read
The short answer
To train a character LoRA: collect 20 to 40 varied, face-correct images of your character, caption each one with a unique trigger word, train on a GPU with around 24 GB of memory (rented by the hour if needed), then test the LoRA on prompts outside the dataset at strengths between 0.6 and 1.0.
The dataset decides the result
A LoRA learns everything the images have in common. If every image is a close-up in soft light, she will only look right in close-ups in soft light. Variety is what makes her usable.
- Framing: close-ups, half body and full body
- Angles: front, three-quarter and profile
- Expressions: neutral, smiling, serious
- Light and places: indoor, outdoor, day and night
- Outfits: several, so clothing does not bake into her identity
Curate hard
Remove any image with a wrong face, broken hands in the foreground or heavy artefacts. Twenty excellent images beat sixty mixed ones.
Captions and the trigger word
Give her a unique trigger word that is not a real word, then caption what changes between images (clothes, place, pose) and leave out what should always be true about her. What you describe stays promptable; what you leave out becomes part of her.
Training
Character LoRAs for SDXL and Flux are commonly trained on GPUs with 24 GB or more. Renting one by the hour is usually cheaper than buying for occasional training, as long as the machine is stopped when the run ends.
Save checkpoints along the way. The last one is often overtrained; an earlier epoch frequently generalises better.
Test before you trust it
Prompt her in places, outfits and poses that were not in the dataset, at strengths from 0.6 to 1.0. If she only looks right in dataset-like scenes, the set was too narrow.
Doing it in Musetta
Musetta walks you from one portrait to a finished LoRA: it builds the dataset, rents the training GPU by the hour, downloads the result into your library and stops the machine. The character then uses it in every tool.
Questions
- How many images do I need to train a character LoRA?
- Between 20 and 40 varied, face-correct images is a common range. Quality and variety matter more than count.
- Can I train a LoRA without my own GPU?
- Yes. Renting a 24 GB GPU by the hour works well for occasional training, provided the machine is stopped as soon as the run finishes.
Teaching this to your own audience?Earn 20% of every payment from the people you send, for 6 months.How it works
