What is a character LoRA?
A character LoRA is a small add-on to an image model, trained on many shots of one face, that teaches the model to draw that face the same way every time. Load it with the base model and the character keeps its face across poses, outfits, light and, through image-to-video, clips.
How it works
An image model already knows how to draw faces in general. A LoRA (Low-Rank Adaptation) adds a small set of trained weights on top of it, learned from a set of shots of one face. The base model stays the same; the LoRA nudges it toward that one face whenever it's loaded, usually with a trigger word in the prompt.
Because the LoRA is small, it trains in a single GPU session and is easy to keep, reuse and combine with other settings. That's why it's the usual answer to keeping an AI character consistent.
What training takes
| Step | What happens | In castingcard.ai |
|---|---|---|
| A training set | Varied, good shots of the same face | Planned from one reference photo as a 24-shot set; you keep the shots that look right |
| Training | A GPU learns the face from the set | Runs on your own GPU or a rented Runpod GPU, with the cost shown first |
| Choosing a checkpoint | Earlier and later saves are compared | The studio draws the same test prompts with several checkpoints, scores the likeness and suggests one |
| Refining | Train again on cleaner shots | Optional: the first LoRA draws a fresh set, you keep the best, and it trains again |
What you use it for
- A brand's recurring model across lookbooks and product videos.
- A UGC creator who fronts every ad in a set: AI UGC ads.
- An AI influencer whose feed always shows the same person: AI influencer studio.
Questions
- What does LoRA stand for?
- Low-Rank Adaptation: a way to fine-tune a large model by training a small set of extra weights instead of the whole model. The result is a small file used alongside the base model.
- What's the difference between a character LoRA and a style LoRA?
- A character LoRA learns one subject, such as a face. A style LoRA learns a look, such as a film stock or an illustration style. You can use both at once.
- Do I need to know how to train a LoRA to use castingcard.ai?
- No. The studio plans the training set from one reference photo, trains the LoRA, compares checkpoints and suggests the best one. You pick the shots and the final checkpoint.
- Can I train a LoRA of a real person?
- No. Every character is a fictional adult. Real people, celebrities and lookalikes aren't allowed.
- Do I need my own GPU?
- No. Use a graphics card you already have, or rent one from Runpod by the second on your own Runpod account. The studio sets the GPU up, runs your jobs and shuts it off when idle. A photo takes about 0.1¢ of GPU time and a 5-second clip about 39¢ on a rented RTX 4090.
Cast your first model
Train a character LoRA from one reference photo, in the browser.