Train a character

Train Phosphene on one person’s face — and, optionally, their voice — from your own photos. The result is a character you can put in any LTX shot. Training needs at least 24 GB of memory and runs on LTX-2.3 (the tab downloads what it needs).

The steps

1. Train type — Character (the face, and voice, of one person) or Style (a cinematic look, colour, lighting; experimental).
2. Dataset — drop 15 to 50 photos (up to 500) — PNG, JPG or WEBP, with matching .txt captions if you have them — or a ZIP of paired images and captions. Each preview shows the square crop the trainer sees. How to train a character well covers the mix of shots, light and backgrounds that works.
3. Captions — see below.
4. Trigger word — a rare, letters-only word the model ties to this character. ↻ Suggest makes one. Avoid digits: they split into common tokens.
5. Quality preset — see below.
6. Crop strategy — Center crop for tight portraits, Letterbox to keep wide-shot proportions.
7. Voice (optional) — see below.
8. Check the estimate — Estimated wall time, Peak RAM — and press Train Character.

Captions

Each image needs a caption in the form [VISUAL]: trigger, 50–80 words followed by [TEXT]: None; missing tags are added for you. Auto-caption with Gemma 3 writes one for every image (2–3 seconds each) and overwrites existing captions. A thumbnail marked no cap falls back to a caption made from the trigger word alone.

Captions cannot be edited inside the tab: to change one, upload a corrected .txt (or ZIP). Training asks before it starts if captions are missing or very short.

Presets, and what “identity graded” means

Identity graded means the recipe has actually been measured on real faces and holds them. Only High on a 64 GB+ Mac is. Quick and Medium are for a fast look or a style, not a person.

On Macs under 64 GB, training runs a compact profile (fewer steps, lower rank), and even “High” there is not the graded recipe — the note under the presets says so. Advanced exposes rank, steps, learning rate, resolution and caption strategy.

When a run finishes, Phosphene measures the result. A character that came out weak or carries nothing gets a WEAK or DEAD badge and a banner with advice — usually, train again on High.

Voice

Upload one clean clip — 10 to 25 seconds of a single speaker, MP3, WAV, M4A or FLAC, up to 50 MB — and tick Train voice LoRA. Audio steps: Smoke 100, Standard 250 (default) or Long 500. The button becomes Train Character + Voice.

Using a character in Video

Trained characters land in mlx_models/loras/ and appear in two places:

• The Character mode on the Video tab (LTX): pick the face from the strip. Voiced characters carry a music-note badge. Set strength (0–2, default 1.0), or split it into separate face and voice strengths. No voice keeps the face and drops the speech.
• The LoRA picker, marked Trained, like any other LoRA.

Use the trigger word in the prompt — mrztrn man walking on the beach. A trained character needs the full model: with one active, Quick and Standard are greyed out and the render uses High. The pencil on the strip renames or deletes a character; the trigger word itself cannot change.

Clicking a character chip under Use your trained characters on the Train tab switches to the Video tab in Text mode with that character’s LoRA on at strength 1.0, and fills an empty prompt with its trigger word.