Characters & LoRAs

A character is a reusable identity — a trained LoRA, its trigger words, and reference images, registered under a name. Once registered, you reference the character by name in any generation and FlixML resolves the LoRA that character has for that workflow and injects its trigger word with it. Same character, consistent across every image and video.

Registering a character

Characters carry:

• a LoRA association (the trained weights), named per workflow
• trigger words, injected into the prompt when that character’s LoRA loads
• reference images for consistency

Using a character

Pass the character name (or ID) to a generate call and FlixML handles the LoRA loading and prompt injection for you — you just write the scene.

A LoRA entry names the one workflow it loads in. Generate with a workflow the character has no entry for and no LoRA loads and no trigger word is injected — a trigger only means something to the weights it was trained into, and on its own it is a name at the front of the prompt for the base model to interpret. The likeness is then the character’s base prompt alone.

Training a LoRA

FlixML can train character LoRAs through the API — register a dataset, start training, and monitor checkpoints.

Training runs on a GPU you point it at. Two options:

• Bring your own AI Toolkit server — set AITK_API_URL (and optionally AITK_API_TOKEN, AITK_GPU_IDS) and FlixML drives it.
• Disposable cloud GPU — FlixML can provision a short-lived GPU droplet, install the toolkit, run the training, and tear it down.

Variable	Description
AITK_API_URL	AI Toolkit API URL
AITK_API_TOKEN	AI Toolkit API token (optional)
AITK_GPU_IDS	GPU IDs to use for training (optional)

Full training endpoints and payloads are in the Agent API and SKILL.md.