MODEL IDlightricks:ltx@2.5-pro

live

LTX-2.5 Pro

by LightricksAugust 11, 2026

LTX-2.5 Pro is the higher-capability model in the LTX 2.5 video family, built for production-quality multimodal video creation and transformation. It supports text-to-video, image-to-video, and audio-to-video workflows together with retake, extension, and reframing operations, pairing synchronized audio generation with higher-resolution output and longer-form clip control. It is well suited to commercial video production, concept films, storyboards, speaking-character shots, and other workflows that need stronger quality, broader modality support, and more flexible post-generation iteration than the Fast tier.

Prompting LTX-2.5 Pro

How to write text-to-video prompts for LTX-2.5 Pro: the six-part shot scaffold, directing the action and camera, matching detail to shot scale, and prompting its native audio.

Introduction

LTX-2.5 Pro is Lightricks' production video model, and it reads a prompt the way a director reads a shot brief. You name the shot, the scene, the action, the light, and the sound in one flowing paragraph, and it returns a clip that holds all of them together, synchronized audio included. The clearer the brief, the closer the result lands to a shot a crew set up rather than a moving still.

One text-to-video call at 1080p, with its own generated audio

A cinematic aerial shot at dawn over a wide misty valley, a cluster of colorful hot-air balloons rising slowly through the golden morning haze. The camera cranes gently upward and drifts back, revealing dozens of balloons in reds, oranges, and blues floating above a patchwork of fields and a winding river far below, the low sun flaring through the mist. Warm golden light, a rich cinematic color grade, crisp realistic detail, a strong sense of scale and serenity. The audio is soft and open: the occasional deep whoosh of a balloon burner firing, a gentle high-altitude breeze, distant birdsong, no music.

That clip came from a single call. The prompt set the framing, the subject, the action across the runtime, the light, and the sound of the burners. This guide covers the request shape, the six-part scaffold LTX-2.5 rewards, directing the action, matching prompt detail to shot scale, and the native audio track. Joining several shots into one clip is the multi-shot guide, and driving a still with an audio track is the audio-driven guide.

The request

Every call is a videoInference task, and the only required field is positivePrompt. Everything else has a sensible default.

import { createClient } from '@runware/sdk'

const client = await createClient({ apiKey: process.env.RUNWARE_API_KEY })
await client.connect()

const [result] = await client.run({
  model: 'lightricks:ltx@2.5-pro',
  positivePrompt: 'A cinematic aerial shot at dawn over a wide misty valley, a cluster of colorful hot-air balloons rising slowly through the golden morning haze. The camera cranes gently upward and drifts back to reveal dozens of balloons floating above a patchwork of fields and a winding river far below, the low sun flaring through the mist. Warm golden light, a rich cinematic grade. The audio is the occasional deep whoosh of a balloon burner, a gentle high-altitude breeze, and distant birdsong, no music.',
  width: 1920,
  height: 1080,
  duration: 8,
  settings: {
    audio: true
  }
})

import asyncio
import os

from runware import Runware


async def main():
    async with Runware(api_key=os.environ["RUNWARE_API_KEY"]) as client:
        results = await client.run({
            "model": "lightricks:ltx@2.5-pro",
            "positivePrompt": "A cinematic aerial shot at dawn over a wide misty valley, a cluster of colorful hot-air balloons rising slowly through the golden morning haze. The camera cranes gently upward and drifts back to reveal dozens of balloons floating above a patchwork of fields and a winding river far below, the low sun flaring through the mist. Warm golden light, a rich cinematic grade. The audio is the occasional deep whoosh of a balloon burner, a gentle high-altitude breeze, and distant birdsong, no music.",
            "width": 1920,
            "height": 1080,
            "duration": 8,
            "settings": {
                "audio": True
            }
        })


asyncio.run(main())

curl https://api.runware.ai/v1 \
  -H "Authorization: Bearer $RUNWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[
    {
      "taskType": "videoInference",
      "taskUUID": "b3f8a2c1-5d6e-4790-b1a2-c3d4e5f60718",
      "model": "lightricks:ltx@2.5-pro",
      "positivePrompt": "A cinematic aerial shot at dawn over a wide misty valley, a cluster of colorful hot-air balloons rising slowly through the golden morning haze. The camera cranes gently upward and drifts back to reveal dozens of balloons floating above a patchwork of fields and a winding river far below, the low sun flaring through the mist. Warm golden light, a rich cinematic grade. The audio is the occasional deep whoosh of a balloon burner, a gentle high-altitude breeze, and distant birdsong, no music.",
      "width": 1920,
      "height": 1080,
      "duration": 8,
      "settings": {
        "audio": true
      }
    }
  ]'

runware run lightricks:ltx@2.5-pro \
  positivePrompt="A cinematic aerial shot at dawn over a wide misty valley, a cluster of colorful hot-air balloons rising slowly through the golden morning haze. The camera cranes gently upward and drifts back to reveal dozens of balloons floating above a patchwork of fields and a winding river far below, the low sun flaring through the mist. Warm golden light, a rich cinematic grade. The audio is the occasional deep whoosh of a balloon burner, a gentle high-altitude breeze, and distant birdsong, no music." \
  width=1920 \
  height=1080 \
  duration=8 \
  settings.audio=true

{
  "taskType": "videoInference",
  "taskUUID": "b3f8a2c1-5d6e-4790-b1a2-c3d4e5f60718",
  "model": "lightricks:ltx@2.5-pro",
  "positivePrompt": "A cinematic aerial shot at dawn over a wide misty valley, a cluster of colorful hot-air balloons rising slowly through the golden morning haze. The camera cranes gently upward and drifts back to reveal dozens of balloons floating above a patchwork of fields and a winding river far below, the low sun flaring through the mist. Warm golden light, a rich cinematic grade. The audio is the occasional deep whoosh of a balloon burner, a gentle high-altitude breeze, and distant birdsong, no music.",
  "width": 1920,
  "height": 1080,
  "duration": 8,
  "settings": {
    "audio": true
  }
}

Response

[
  {
    "taskType": "videoInference",
    "taskUUID": "b3f8a2c1-5d6e-4790-b1a2-c3d4e5f60718",
    "videoUUID": "7c1d2e3f-4a5b-6789-0abc-def123456789",
    "videoURL": "https://vm.runware.ai/video/os/a14d18/ws/2/vi/7c1d2e3f-4a5b-6789-0abc-def123456789.mp4"
  }
]

The essentials:

positivePrompt is the only required field, up to 5,000 characters. LTX-2.5 has strong prompt adherence, so a focused brief often beats an exhaustive one. Add detail where the shot needs it and stop there.
width and height are a fixed pair: 1280 × 720 or 1920 × 1080 for landscape, 720 × 1280 or 1080 × 1920 for portrait. Pro renders up to 1080p.
duration is 6, 8, or 10 seconds, or "auto" to let the model read the length from the action you describe.
fps is 24, 25, or 50. 24 reads as cinematic, 25 is broadcast standard, and 50 is smoother motion.
settings.audio set to true adds a synchronized audio track generated in the same pass. Describe the sound in the prompt and the model scores the scene to match.
Video delivery is async: the first response acknowledges the task, and the finished clip arrives by polling or a webhook.

LTX-2.5 Pro renders up to 1080p in clips of 6, 8, or 10 seconds. For 2K or 4K output, clips up to 20 seconds, or 48 fps, reach for LTX-2.5 Fast, which shares this exact prompt structure.

Structuring the prompt

LTX-2.5 rewards a prompt that separates six concerns: the shot, the scene, the action, the character, the camera, and the audio. Each becomes a directive the model tries to hold, so naming all six keeps the decisions with you instead of the model's defaults. Write them as one flowing paragraph in the present tense, not a bulleted list.

a wide shot easing into a medium, late-afternoon sun through floor-to-ceiling windows, a modern open-plan living room with pale oak floors, a linen sofa, and green plants, she walks in from the hallway holding a mug, crosses to the window, and stops to look out over the city as the light catches the steam, a woman in a cream knit sweater, the camera pushes in gently from the doorway and follows her toward the glass, quiet footsteps on the wood floor, a faint city hum through the window, no music

ShotSceneActionCharacterCameraAudio

All six layers directed in one prompt

A warm real-estate lifestyle scene, a wide shot easing into a medium. Late-afternoon sun pours through floor-to-ceiling windows into a modern open-plan living room with pale oak floors, a low linen sofa, and a scatter of green plants. A woman in a cream knit sweater walks slowly in from the hallway holding a mug, crosses to the window, and stops to look out over the city as the light catches the rising steam. The camera pushes in gently from the doorway and follows her toward the glass. Soft golden-hour light, a warm natural color grade, shallow depth of field, realistic materials and reflections. Gentle ambience: quiet footsteps on the wood floor, a faint city hum through the window, no music.

Each layer earns its place. The character clause is distinct from the action: describe who the person is once (age, hair, wardrobe), then let the action clause carry what they do. Emotion reads better as a physical cue than a label, so write "her shoulders drop and she exhales" rather than "she looks relieved." Keep one coherent light logic per shot, since mixed sources confuse the result, and lead with the core shot before you layer in detail.

Two things LTX-2.5 still handles unevenly. In-frame text renders better than earlier versions, but exact spelling and frame-to-frame consistency aren't guaranteed, so keep any text short, verify it across the clip, and composite critical titles or logos in post. Highly chaotic motion can smear, so plausible movement is more reliable than a crowded action, though everyday motion like a walk or a dance is fine.

Directing the action

Motion comes from two places worth prompting separately: what the camera does, and what the subject does. Keep them in separate clauses so the model knows which is which. In the clip below the camera is locked and every bit of movement belongs to the subject.

A locked camera, all motion carried by the subject

A high-energy fitness scene, a medium shot on a locked, static camera. A lean boxer in black training gear works through a fast combination in a dim industrial gym, hard shafts of window light cutting through drifting chalk dust. She snaps out a jab and a cross, rolls her head under an imaginary counter, and resets her stance with her breath ragged. The camera never moves; every bit of motion is her footwork, her gloves, and the dust turning in the light. A moody high-contrast grade, deep shadows, crisp detail. The audio is her sharp exhale on each punch, the scuff of her shoes on the concrete, and a low industrial room tone, no music.

Write the subject's action as a natural sequence that flows from beginning to end, in present tense: she snaps out a combination, rolls under a counter, resets. Naming the beats in order lets the model pace the runtime to fit them. Hold to one dominant action arc per clip, since two arcs competing for the same seconds lose fidelity on both. When you want the camera to move as well, say so in its own clause, and the camera guide covers the full vocabulary and the cameraMovement presets.

Matching detail to shot scale

How much to write depends on how tight the frame is. A wide shot reads from big shapes and light, so it needs less surface detail. A close-up fills the frame with texture, so every material has to be named or the model invents it. The same lunch, framed wide and then macro, shows the difference: the wide prompt sets the table and the light, while the close-up prompt has to specify the char, the butter, and the steam.

Wide: shapes, layout, and light

A wide overhead food shot of a summer lunch laid out on a sunlit terrace table: a platter of grilled vegetables, a board of torn bread, a bowl of bright salad, and tall glasses of iced tea on a rumpled linen cloth. The camera drifts slowly straight down over the whole table as a hand reaches in to set a final dish in place. Bright natural daylight, a fresh editorial food grade, clean realistic detail. Audio: light outdoor ambience, the soft clink of a plate settling on the table, distant birdsong, no music.

Close: named texture, char, and steam

An extreme macro close-up from the same lunch: a fork lifts a single piece of grilled corn slowly up out of the frame, melted butter glistening and dripping along the charred kernels, tiny flecks of herb and black char in sharp focus, a thin wisp of steam curling off it. The camera holds in a slow tilt that follows the fork upward. Warm natural light, a shallow macro depth of field, pristine food detail. Audio: a faint sizzle, a soft outdoor ambience, no music.

The rule of thumb: the tighter the shot, the more concrete the nouns. Reach for a wide when the story is the whole scene, and a close-up when it is a single detail, then load the prompt to match.

Describing the sound

Audio is part of the prompt, not an afterthought. Turn it on with settings.audio, then name what you want to hear in a plain sentence, and LTX-2.5 builds a track locked to the picture. The clip below is written around its sound, the boom of each wave against the rocks and the hiss of the foam draining back.

A soundscape built in the prompt, the boom and hiss of surf on the rocks

A dramatic coastal seascape at dawn, a wide shot on a rugged rocky headland. Powerful ocean waves surge in and crash against the dark jagged rocks, exploding into white spray and sheets of foam that rush over the stone before draining back into the churning sea, a fine mist drifting in the low golden light. The camera holds steady on the rocks as the waves roll in and break. Cinematic cool-and-warm grade, crisp realistic water detail, a strong sense of scale. The audio leads the shot: the deep boom of each wave hitting the rocks, the hiss and rush of the foam draining back, the constant roll of the surf, and distant gull calls, no music.

Name the ambient bed and the specific effects tied to the action, and say whether there is music. Adding "no music" keeps the mix to the room and the action when you plan to drop the clip onto a soundtrack of your own. Spoken dialogue and singing have their own patterns in the native audio guide.

Sizes, duration, and frame rate

The same prompt structure carries any shape. Pass a landscape pair for a hero or a broadcast cut, and a portrait pair (1080 × 1920) for stories and reels, where the model recomposes for the tall frame rather than cropping a wide one.

A 9:16 portrait render for social

A vertical fashion film shot for social, a full-length portrait framing. A model in a flowing emerald silk dress walks slowly toward the camera down a sunlit marble corridor, the fabric rippling and catching the light with every step, then stops and turns her shoulder to the lens. A slow steady push-in holds her head to toe in the tall frame. Bright directional daylight, an elegant editorial color grade, shallow depth of field. The audio is the soft click of her heels echoing down the corridor, the light rustle of silk, and quiet ambient air, no music.

Pick duration by the beats the shot has to fit: 6 seconds for one clean action, 8 or 10 for a short sequence that develops. Set it to "auto" and the model reads the length from the action you describe, so a one-line moment stays short and a fuller sequence runs longer. fps sets the motion feel, 24 for a cinematic look and 50 for smoother sport and action.

Tips

Name all six layers. Shot, scene, action, character, camera, and audio. Dropping any one hands that decision to the model, and the audio layer is the one most often forgotten.
Write one flowing paragraph in present tense. LTX reads a continuous scene description, not a shot list. Keep the action moving from beginning to end.
Separate camera motion from subject motion. Put the camera move in its own clause and the subject's action in another, so the model knows which to animate.
Match detail to shot scale. Close-ups need named texture and material. Wide shots read from shapes and light. Load the prompt for the frame you asked for.
Show emotion with physical cues. "She exhales and her shoulders drop" directs the performance where "she looks relieved" leaves it to the model.
Write the sound inline. Turn on settings.audio, name the ambient bed and the effects tied to the action, and add "no music" when a scored track would clash with your own.
Keep in-frame text short. Spelling and frame-to-frame consistency aren't guaranteed, so keep any text brief, verify it across the clip, and composite critical titles or logos in post.