# LTX-2.5 prompting guide: scenes, sound and dialogue

> Write LTX-2.5 prompts for a live stream with sound: one audio-visual brief per ten-second window, dialogue in quotes, continuity anchors and when to cut.

Source: https://windflow.online/guides/ltx-2-5-prompting-guide · Updated 2026-11-03 · Windflow Labs

[LTX-2.5](https://windflow.online/models/ltx-2-5) generates picture and sound together, one window of about ten seconds at a time. A good prompt is a small screenplay for that window: what the camera sees, what happens, and what the scene sounds like. This guide covers how to write one, how to chain windows into a stream, and what the model handles badly.

For the API (tokens, the WebRTC room, audio playback), see [how to use the LTX-2.5 API](https://windflow.online/guides/ltx-2-5-api).

## Write one audio-visual brief per window

Lightricks' own prompting advice for LTX-2 comes down to a few habits, and they apply here:

- **Write one flowing paragraph**, in the present tense, that reads from the start of the window to the end. Expect four to eight sentences.
- **Open with the shot**: framing and genre in film language, such as "a tight close-up", "a wide establishing shot", "handheld, documentary style".
- **Set the scene**: light, colour, textures and atmosphere.
- **Describe the action in order**, as it unfolds.
- **Describe characters by what can be seen**: age, hair, clothing, distinguishing details. Show emotion through posture and expression rather than words like "sad".
- **Say how the camera moves**, and what is in frame after the move.
- **Describe the sound**: ambience, music, effects, and any speech.

| Weak | Strong |
|---|---|
| A busy market. | A handheld walk through a rainy night market. Vendors call out from both sides under glowing paper lanterns, and water drips steadily from canvas awnings. The camera weaves past a woman in a yellow raincoat choosing oranges. Traffic hisses on the wet street beyond the stalls. |

Match the detail to the shot: a close-up needs precise detail about a face or an object, a wide shot needs less.

## Write dialogue and sound into the prompt

Put spoken lines in quotation marks, and say who speaks and how: `The man leans toward the camera and whispers, "Today I finally get it right."` Name the language or accent if it matters.

For sound, name concrete sources that belong in the picture: footsteps on gravel, rain on a tin roof, an espresso machine, distant traffic, room tone. One or two clear sources work better than a long list. Audio is generated together with the picture, so the more concretely you describe it, the closer it follows your intent.

## Chain windows into a stream

Nothing steers inside a window. `set_prompt` changes what the **next** window generates, so think of each prompt as the next beat.

- **Continue the scene** with `set_prompt`. The next window continues from the tail of the previous one, so identity, wardrobe, room and lighting carry across. The camera can still step at the boundary. Repeat the stable details in the new prompt, and change one major thing per window.
- **Cut to a new scene** with `scene_cut`. It drops the previous tail and starts an independent window. Describe the whole new scene, because nothing carries over.
- **Swap the reference still and prompt together** with `set_conditioning` when the next window should re-anchor on a new image.

| Intent | Command | What carries over |
|---|---|---|
| Continue the same scene | `set_prompt` | The previous window's visual tail |
| Re-anchor on a new still | `set_conditioning` | The new still |
| Start an unrelated scene | `scene_cut` | Nothing |

A new cast, location, time of day and lens all at once is a cut, even if you send it with `set_prompt`. Send it as `scene_cut` and the boundary will look intentional.

## Open on a reference still

A reference still anchors the opening composition and identity. Write the prompt as what happens next in that image, not as a different picture: if the still shows a woman at a window in daylight, `She turns from the window and smiles as rain starts` extends it, while a prompt about a night street contradicts it. Later windows continue from generated frames rather than returning to the still.

## Write for what LTX-2.5 draws well

Lightricks' guidance for LTX-2 points the same way for the realtime profile:

- **Keep text out of the scene.** Describe the shop, not the words on its sign; leave out labels and brand names.
- **Favour natural, readable motion.** Walking, gestures, dance and camera moves come through clearly; keep juggling and acrobatics out of the brief.
- **Keep the cast small.** One or two characters with a clear action read best. Layer in more detail over later windows.
- **Keep the light consistent.** Choose light sources that belong together, unless the scene calls for the contrast.
- **Show emotion.** Describe posture, gesture and expression rather than words like "confused".

Start simple, watch a window, and add one instruction at a time.

## Tune the result

| Symptom | Likely cause | What to change |
|---|---|---|
| The shot shifts at a boundary | Too much changed between windows | Repeat the stable details; change one thing per window |
| The sound is not what you wanted | Sound direction was vague or crowded | Name one or two concrete sources that are in the picture |
| Part of the prompt does not show up | Too many actions or characters | Cut the prompt down and layer details back in over later windows |
| Signs or labels look off | Text in the scene | Describe the scene without written text |
| A change appears a little later | It waited for the running window | Commands land at the next window boundary |

## FAQ

### How long should an LTX-2.5 prompt be?

Four to eight sentences in one paragraph, covering the shot, the scene, the action in order, the characters, the camera and the sound.

### Can LTX-2.5 generate speech?

Yes. Write the line in quotation marks and say who speaks and how.

### How do I keep a character consistent across windows?

Repeat their visible details (age, hair, clothing) in every prompt and use `set_prompt` rather than `scene_cut`, so each window continues from the previous one's tail. A reference still helps anchor the opening.