Train a character LoRA

Reference-mode is the cheap path: the renderer reads your reference images and prompts the image model to match. It works for 80% of cases. The other 20% - identity drift on tight close-ups, motion inconsistency on long clips, brand-critical mascots that have to be exact every time - is what character LoRAs are for.

This guide covers the image LoRA and video LoRA training flows. Studio tier and up.

When to train

Train a LoRA when one of these is true:

• The character has to be exact. Brand mascots, licensed IP, real-person digital twins.
• You render the same character hundreds of times. Train-once amortizes across every video.
• Reference-mode is drifting. You see it in long clips, in close-ups, in unusual angles.

The flow

From the Character detail page at /app/characters/{slug}:

1. Hit Train under the Image LoRA rail (or Video LoRA, separately).
2. The trainer pulls the character's reference images and clips. Image LoRAs need 8 to 20 images; video LoRAs need 3 to 8 clips of 4 to 8 seconds each.
3. Status moves through queued, training, ready. Image LoRAs take 20 to 60 minutes; video LoRAs take 1 to 4 hours.
4. When ready, switch the character's image_mode or video_mode dial from ref to lora. Subsequent renders use the trained model.

Costs

Each training job consumes character LoRA credits. The exact cost depends on the provider; the credit deduction shows in your usage dashboard. Failed jobs do not consume credits.

Per-render override

You don't have to commit to LoRA for every render. The per-video override on the Edit page lets you flip Ref vs LoRA per cut. Reach for LoRA for hero pieces, stay on Ref for quick experiments.

Bring it to VideoCue

Studio tier unlocks the image LoRA rail; Portfolio tier unlocks both image and video. Open any character at /app/characters, scroll to the Training section, and hit Train. The webhook updates the model status the moment training finishes.