LTX 2.5: Multi-Shot Video, One Pass, Real Sound

Here's how most AI video "scenes" actually get made. You generate a wide shot. Then you generate a close-up, praying the woman in it is the same woman. Then a reverse angle, where her coat has quietly changed colour and the rain has stopped for no reason. Then you open an editor, drop in a stock ambience track, and spend an hour hiding the seams.

Three clips stitched together is not a scene. It's a ransom note.

LTX 2.5 from Lightricks goes after that exact problem. Its headline trick is native multi-shot generation: several connected shots, with real cuts between them, produced in a single pass, with sound generated alongside the picture. Lightricks says those shots hold character, environment, lighting and voice across the cuts. That last word matters more than it looks. Voice drift is the continuity error nobody talks about until they hear it.

What LTX 2.5 Actually Is

LTX 2.5 is the newest open-weight video model from Lightricks, the team behind the LTX family. The Hugging Face card describes it as an "open world model with open weights, built for local execution and fine-tuning". It generates video and synchronized audio together, from text, from an image, or (on each::labs) from a clip of sound.

The upgrade list from its predecessor, LTX 2.3, is long, but four changes do the heavy lifting. A new diffusion-based video decoder replaces the old reconstruction stage, which Lightricks says brings cleaner motion and better faces, textures and on-screen text. A Gemma 4 12B text encoder, paired with a custom prompt enhancer, is there so dense prompts (several characters, a camera move, a lighting change, a line of dialogue) don't get flattened into mush. Automatic duration prediction lets the model choose a clip length that fits the action you described. And there's multi-shot, which earlier LTX versions simply didn't do: they gave you one continuous take.

Native 4K HDR support rounds it out for teams who grade and composite. You don't need that to get value from it on each::labs, but it tells you who the model was built for: people making footage that has to survive an edit.

One Prompt, Several Shots, Real Cuts

Multi-shot is the reason to pay attention, so it's worth being precise about what it is. There's no "number of shots" setting in LTX 2.5. You write the cuts into the prompt, in prose, and the model edits the scene for you. A wide establishing shot, a hard cut to a medium, a cut to a close-up on a detail. One generation, one file, one continuous soundtrack running underneath.

Why does that beat generating three clips and cutting them yourself? Because the model is solving the whole sequence at once. When the same generation contains the wide and the close-up, there's no separate roll of the dice for each angle. The character in shot three is the character from shot one because they were never separate requests. Lightricks' claim is that identity, environment, lighting, visual style and voice all carry across the cut, and that's the list you'd actually want if you were the person stuck fixing continuity in post.

Restraint pays here. Lightricks' own prompting guide suggests two to four shots in a single generation, each with a clear job. Pile on more cuts and every shot gets less time to establish itself, which is how you end up with a montage instead of a scene. And sometimes you don't want a cut at all: the same guide steers you toward one continuous take for unbroken camera moves and lip-synced dialogue in a single framing.

Sound Comes Out of the Same Pass

Every text and image variant of LTX 2.5 on each::labs ships with audio switched on by default. Ambience, effects, music, speech: the model generates them together with the frames, so the door slam lands when the door slams. You can turn audio off with a single flag when you plan to score the piece yourself.

The interesting part isn't that audio exists. Plenty of current models do that now. It's how audio interacts with multi-shot. Lightricks' guide tells you to state, at each cut, whether the audio continues or changes. So a synth score can run unbroken across three angles while the rain gets louder in the close-up, or a room tone can drop out hard on a cut to an exterior. That's editing language, and the model listens to it.

Dialogue goes in quotation marks, with the speaker named and, if it matters, the language or accent. A held, lip-synced line is a single-take job, so don't cut away mid-sentence.

Six Models, Three Ways In

On each::labs, LTX 2.5 comes as six models: three input routes, each in a Fast and a Pro tier. Which route you pick depends on what you already have.

Text to Video Pro and Text to Video Fast take a prompt of up to 5,000 characters and nothing else. That's plenty of room for a three-shot scene with dialogue and sound cues. You choose 16:9 or 9:16, a duration, a frame rate and a resolution, and you can add an optional camera move: dolly in or out, dolly left or right, jib up or down, a focus shift, or a locked-off static frame.

Image to Video Pro and Image to Video Fast start from a picture you supply. Both accept an optional end image, in which case the model generates the transition between your first and last frame. That's the control you want when a client has approved the opening composition and the closing one, and you just need the motion in between. Aspect ratio defaults to auto, so the output follows your source image.

The third route is the unusual one, and it gets its own section below.

Fast Isn't the Draft Tier Here

Most model lineups follow the same script: Fast is the sketchbook, Pro is the real thing, and Pro can do everything Fast can and more. LTX 2.5 doesn't follow that script, and it's worth reading the parameters before you assume.

On the text and image routes, the Fast tier is the one with the bigger envelope. It renders from 6 up to 20 seconds and goes up to 1440p and 2160p, with 48 fps on the menu alongside 24, 25 and 50. Pro sticks to 6, 8 or 10 seconds at 720p or 1080p, with 24, 25 or 50 fps. Image to Video Pro adds an auto duration option, letting the model's duration predictor pick the length for you.

So how do you choose? each::labs positions Fast for iteration and Pro for the finished take, and that's a sensible default for clips that fit inside ten seconds at 1080p. But if the deliverable is a 4K hero shot, or a sequence that needs 15 seconds to breathe, Fast is the only tier that can produce it. Pick by the spec of the deliverable first, then by phase.

Let the Soundtrack Direct

Most video models treat sound as something that happens to the picture. Audio to Video Pro flips that. You hand it an audio file, and it builds the visuals around it.

The inputs are short and focused. An audio URL is required: 2 to 10 seconds on Pro, 2 to 20 seconds on Audio to Video Fast. You can add an image to use as the first frame, and you need a prompt if you don't supply one. A guidance scale from 1 to 50 sets how tightly the model follows your text, and the aspect ratio can be auto, 16:9 or 9:16.

This is the route for work where the sound already exists and isn't up for negotiation. A recorded line plus a portrait becomes a talking shot. A brand's sonic logo becomes a five-second sting that moves on the beat. For music videos, Lightricks' own tutorial splits the track into sections, generates a shot for each, then assembles. If your edits usually start with "the client sent the voiceover", start here.

Write the Cut, Don't Hope for It

LTX 2.5 rewards a specific prompt shape, and it's not a shot list. Lightricks recommends writing the whole scene as one chronological paragraph, in present tense, and describing each cut in prose: name the transition (hard cut, match cut, dissolve), re-establish the new framing, keep the recurring subject identifiable, and say what the audio does. Here's a three-shot example for Text to Video:

A wide shot of a rain-soaked night market, neon signs reflecting in puddles, a woman in her thirties with a cropped silver bob and a yellow raincoat weaving between stalls, a low synth score and the hiss of rain. Hard cut to a medium shot of the same woman stopping at a noodle stall, steam rising past her face, the synth score continuing under the clatter of bowls. Hard cut to a close-up of her hands counting out coins on the wet counter, rain louder now, the score fading to a single held note.

Notice what's doing the work. Each shot has a job (place, person, detail). The character's description is concrete enough to survive two cuts: age, hair, one bold colour. And the audio has its own arc across the sequence.

For dialogue, keep it to one framing and give the model physical cues rather than labels. "She's nervous" gives it nothing to animate. Fidgeting hands do.

A medium close-up of an older fisherman with a grey beard and a wool cap, sitting on an overturned crate at dawn, harbour gulls calling in the distance. He rubs his thumb along a frayed rope, looks up at the camera and says, in a low gravelly voice, "Forty years, and the sea still surprises me." He pauses, the gulls fill the silence, and he smiles.

That written pause is deliberate. Lightricks notes that the duration predictor won't invent beats you didn't write, so if you want silence, write it. Prompts written for another model need rewriting into this paragraph shape, not pasting. If you want a deeper bench of starting points, our AI video prompt examples translate well once you add the cuts and sound cues.

Open Weights Make It Something You Can Build On

The open-weight release matters even if you never download a checkpoint. Published weights are a fixed reference point: the community can study them, and teams with special requirements can fine-tune on their own footage using the base checkpoint Lightricks ships for exactly that. If you're wiring generated video into a product feature, you want the behaviour you tuned against to still be there next quarter. The full picture, including HDR and alpha matte tooling, is on Lightricks' LTX 2.5 page.

Running LTX 2.5 on each::labs

All six variants use the same prediction endpoint, so switching routes is a model string and an input shape. A multi-shot text call looks like this:

curl -X POST \
  -H "Authorization: Bearer $EACHLABS_API_KEY" \
  -H "Content-Type: application/json" \
  --data '{
    "model": "ltx-2-5-text-to-video-fast",
    "input": {
      "prompt": "A wide shot of a rain-soaked night market... Hard cut to a medium shot of the same woman... Hard cut to a close-up of her hands...",
      "duration": 10,
      "resolution": "1080p",
      "aspect_ratio": "16:9",
      "fps": 25,
      "generate_audio": true
    },
    "webhook_url": ""
  }' \
  https://api.eachlabs.ai/v1/prediction/

Point it at ltx-2-5-image-to-video-pro and add image_url (plus end_image_url if you want a first-to-last-frame transition). Point it at ltx-2-5-audio-to-video-fast and pass audio_url, with an optional image_url as the first frame. The response is a prediction ID you poll or receive by webhook, so build for that from day one.

If you're weighing LTX 2.5 against the rest of the field, our ranking of the best AI video generation models in 2026 sorts them by job.

Frequently Asked Questions

How does LTX 2.5 multi-shot generation work?

Through the prompt, not a setting. You describe the scene as one chronological paragraph and write each cut into it ("hard cut to a close-up of..."), and the model returns a single clip with those shots edited together and audio running across them. Lightricks suggests two to four shots in one generation, each with a clear purpose.

Should I use LTX 2.5 Fast or Pro?

Start from the deliverable. Anything longer than 10 seconds or sharper than 1080p needs Fast, which reaches 20 seconds and 2160p. Within those limits, each::labs positions Fast for iteration and Pro for the finished take, so explore on Fast and render the keeper on Pro.

Can LTX 2.5 make a video from my own audio?

Yes, and that's the route that sets this family apart on each::labs. Audio to Video takes a clip of 2 to 10 seconds on Pro or 2 to 20 seconds on Fast, optionally with an image as the opening frame, and builds visuals timed to the sound. Think dialogue, a music cue or any sound-led scene where the audio is already locked.

Does LTX 2.5 generate audio automatically?

On the text and image routes, audio is on by default and generated in the same pass as the picture: ambience, effects, music and quoted dialogue. Switch generate_audio off when you're scoring the piece yourself.