> ## Documentation Index
> Fetch the complete documentation index at: https://docs.reactor.inc/llms.txt
> Use this file to discover all available pages before exploring further.

# LTX prompt guide

> How to choose an avatar photo and write scripts and scene prompts that render well.

LTX renders best from a single well-lit, forward-facing photo, a script with natural punctuation,
and a scene prompt that describes the voice, staging, and delivery. This page covers what each input
controls and where people commonly go wrong.

## What it renders best

* **A single well-lit person facing the camera.** Identity lock across the whole take is the model's
  superpower, and this is the photo that gives it the most to anchor to.
* **Natural speech at 110–160 WPM.** Every persona in the launch demo sits in this range; the pace
  accepts 80 to 220, but the top of the range sounds rushed and can smear lip closures.
* **Voice and delivery driven by the scene prompt.** The model generates audio and video together
  and conditions both on the prompt, so the prompt is where the voice is cast: register, grain,
  accent, breath, cadence, and dynamics.

## Before & after

**Casting the voice**

<div className="rea-example rea-example-do">
  `"A single person speaks directly to the camera. His voice is a low, salt-roughened baritone with real grain in it, chesty and open, worn down by wind and shouting across a deck. A rolling West Country lilt, consonants bitten off short, vowels stretched wide."`

  A casting note: register, grain, accent, and cadence, describing a speaker rather than issuing an
  instruction.
</div>

<div className="rea-example rea-example-dont">
  `"in a rough, cheerful sailor's voice"` or `"make him sound deeper"`

  An adjective label names a type without pinning a read, and two prompts written this way land on the
  same generic voice.
</div>

**Controlling staging vs. identity**

<div className="rea-example rea-example-do">
  `"news anchor at a desk, professional tone"`: staging and delivery, which the scene prompt owns.
</div>

<div className="rea-example rea-example-dont">
  `"change the shirt to red"`: identity and wardrobe are anchored to the photo. Text loses this fight
  every time.
</div>

## The photo is the ceiling

There is no prompt that recovers a bad photo. A single well-lit, forward-facing photo anchors
identity for the entire take and gives the strongest identity lock and the crispest lip-sync. A
low-quality, occluded, or profile photo degrades everything downstream. A single-subject photo is
strongly recommended.

## What each input controls

| Input                             | Controls                                       | Does **not** control        |
| --------------------------------- | ---------------------------------------------- | --------------------------- |
| Avatar photo (`set_avatar_image`) | Identity, appearance, wardrobe                 | Voice, delivery, script     |
| Script (`set_script`)             | The words spoken, and (via punctuation) pacing | Voice timbre, scene, camera |
| Scene prompt (`set_prompt`)       | Voice, staging, lighting, mood, delivery       | Identity, wardrobe          |

<Note>
  Treat these as three independent channels. A prompt that tries to reach across channels will be
  ignored on that axis.
</Note>

## Writing scene prompts

* Describe the voice, staging, lighting, and tone, not the subject's appearance.
* Cast the voice the way a director casts a part: register, grain, accent, breath, cadence, and
  dynamics. Describe the speaker in the third person rather than instructing the model.
* Delivery energy follows the scene prompt: an energetic prompt changes pacing and expression even
  with the same script.
* Keep the prompt under 800 characters, the limit `set_prompt` accepts.
* Send the prompt once and leave it alone. Resending `set_prompt` mid-conversation resets the voice,
  so a character that has been speaking will come back sounding like someone else. Pin `set_seed` to
  a fixed number for the same reason.