Every character in VideoCue has two identity dials: image mode and video mode. Each can be set to Ref or LoRA. The choice maps to a cost-vs-fidelity trade. This guide is the short version of when to pick which.
The trade
Ref-mode. The renderer reads your reference images and prompts the image model to match identity. Cheap, instant, no training step. Works well for most characters but allows some drift across long sequences or unusual angles.
LoRA mode. A per-character fine-tuned model. Locks identity at the model weights, so the character renders identical across every cut. Costs training time and credits. Studio and Portfolio tiers.
Quick decision table
| Situation | Pick | |---|---| | New character, testing the concept | Ref | | Brand mascot, has to be exact | LoRA | | Licensed IP or real person digital twin | LoRA | | Animal pet, casual use | Ref | | Recurring presenter, hundreds of videos | LoRA | | One-off cameo character | Ref | | Identity is drifting on long clips | LoRA | | Free or Creator tier | Ref (LoRA gated to Studio+) |
Per-rail dial
Image mode and video mode are separate. It's common to start with image LoRA and stay on video Ref - image identity drifts more than motion identity, and video LoRA training is the more expensive of the two. Train the image rail first, evaluate, then decide whether the video rail is worth the spend.
Per-render override
You don't have to commit. Every video Edit page has an image_mode_override and video_mode_override selector that overrides the character defaults for that specific render. Use Ref for quick test cuts, LoRA for hero pieces.
Bring it to VideoCue
Pick Ref for new characters; switch to LoRA when you see drift on cuts you actually care about. The pricing rows on /pricing show which tiers unlock LoRA training.