Technical Explainer

Character Training with LoRA — How Twin Lock Works

Upload 20 photos. Wait 30 minutes. Get an AI character model that keeps your face identical across every image and video generation. Here's how it works under the hood.

What is LoRA?

LoRA (Low-Rank Adaptation) is a lightweight fine-tuning technique. Instead of retraining an entire AI model (which costs thousands of dollars and takes days), LoRA trains a small "adapter" — a few megabytes of weights — that sits on top of an existing model and teaches it a specific concept.

In Twin Studio's case, the concept is you. Your face, your wardrobe, your body type, your style. The result is a model called Atom X — your personal AI character.

The Twin Lock pipeline

1. Upload 20 photos — diverse angles, good lighting, different outfits, different expressions
2. Automatic preprocessing — images are cropped, aligned, and normalized
3. LoRA training — the adapter learns your identity features (face shape, skin tone, hair, build)
4. Atom X model created — the trained LoRA is packaged as your personal character model
5. Ready to generate — use Atom X for any image, or feed Atom X images as first frames for video

Why 20 photos?

20 photos gives the LoRA enough variety to learn your face from different angles and lighting conditions. Fewer than 15 photos risks overfitting (the model memorizes specific photos instead of learning your general appearance). More than 30 has diminishing returns.

Tips for best results

• Diverse angles — front, 3/4, profile, slightly above, slightly below. Don't upload 20 selfies from the same angle.
• Good lighting — natural light is best. Avoid harsh shadows or overexposed shots.
• Different outfits — this teaches the model that your identity is separate from your clothing. You can change outfits in generation.
• Different expressions — smiling, neutral, serious. The model needs to know your face moves.
• Consistent person — all 20 photos must be of the same person. No group shots.
• High resolution — at least 512x512 per photo. Phone photos are fine.

Using Atom X for video

Atom X is primarily an image model. For character-consistent video, you:

1. Generate a hero frame with Atom X (your face in the scene)
2. Feed that image as the first frame into a video model (Kling 2.6, Atom V Max, or VEO 3.1)
3. The video model animates the frame while preserving your identity

For maximum identity lock across a full film, use Atom X for every hero frame, then animate each one. Your face stays consistent across all shots — even if the scene, lighting, and wardrobe change completely.