LTX 2.5 Multishot: How It Works and How to Use It

Before LTX-2.5, generating a multi-scene video meant generating scenes separately and hoping they matched. Different runs drift: the character's face shifts, the lighting changes, the voice no longer sounds like the same person. Then you stitch the shots together and the mismatch is exactly at every cut — where the audience is looking.

LTX 2.5 multishot removes that pipeline. The model card describes it as native multishot generation: multiple connected shots in a single pass that hold character identity, environment, lighting, voice, and visual style across cuts. One generation produces the whole sequence, continuity included.

This guide covers what multishot actually is, what it does not do, and how to prompt it — based on the official LTX-2.5 model card, the ltx.io LTX-2.5 page, and the ComfyUI LTX-2.5 tutorial. Checked August 13, 2026.

What multishot actually is

"Shot" here means a continuous camera take — what editors call a shot before cutting to the next angle. Multishot means the model produces several of these, cut together, in one generation.

The official description names the five things that persist across cuts:

1. Character identity — the same person stays the same person between shots
2. Environment — the room or location stays consistent
3. Lighting — the light direction and mood do not jump
4. Voice — dialogue audio stays the same voice
5. Visual style — the look of the footage stays consistent

Previous versions produced a single continuous shot: one camera, no cuts, no scene changes. Multishot is the structural change in 2.5 that lets you generate "two people enter the room, then a close-up of the woman speaking, then the man walks out" as one connected output instead of three separate generations stitched by hand.

What it does not do

Being precise about limits is where this gets useful:

• Multishot is not an edit tool. It generates connected shots. Reordering, replacing, or re-cutting individual shots after the fact is still your job — and that is where LTX-2.5's retake/region pipelines (inherited from the LTX-2 family) come in.
• Consistency is per-generation, not per-identity. If you generate the same character in two separate runs, nothing promises the two runs match. Identity persistence is a within-generation property.
• It is a claim about structure, not a quality guarantee. The vendor describes what the feature is designed to hold across cuts; how well it holds on a given prompt is something you verify on your own clips. The model card's own limitations section notes the model "may fail to generate videos that match the prompt perfectly."

How multishot fits the rest of LTX-2.5

Multishot is one of six headline changes in the official model card; the others are diffusion fidelity rendering (compute allocated by scene complexity), a new diffusion video decoder (sharper faces, textures, and on-screen text), the Gemma 4 12B text encoder, a prompt enhancer, and a duration predictor. The pieces that matter most for multishot prompting are:

• Gemma 4 12B encoder — the official reason complex prompts with multiple subjects, actions, and camera moves survive the encoding stage without details being dropped. Multishot prompts are complex by definition, so this encoder is effectively a prerequisite for reliable multishot output.
• Prompt enhancer — expands a short prompt into richer cinematic instructions at minimal extra compute. For multishot, an enhancer that expands "the same woman, two shots" into detailed per-shot descriptions is more useful than a generic prompt prettifier.
• Auto duration — an optional duration head predicts the clip length from the prompt. For multi-scene prompts, the implied length is longer than a single-shot prompt; the duration head is what lets you omit --num-frames and get a sensible length.

Test multishot before you build a workflow around it. Generate a two-shot sequence in your browser at ltx23.app — a character in one location, then a close-up of that character in the same light — and see how well identity holds on your own prompt style.

How to prompt multishot

The official prompting guidance for the LTX-2 family (from the LTX-2 README) applies, with multishot-specific additions. The base rules: a detailed, chronological description of actions in a single flowing paragraph, under 200 words, literal and precise, written like a cinematographer describing a shot list.

For multishot, structure your prompt as a shot list:

1. Introduce identity and environment once, then reference them. Describe the character's appearance and the location in the first shot. In later shots, refer to "the same man" / "the same room" instead of re-describing — and note that Lightricks' own prompting doc builds prompts this way for consistency across the sequence.

2. Mark the cuts explicitly. Separate shots with clear transition language: "Cut to a close-up of the same woman speaking" / "The camera then moves to…". If the model has no shot boundary to follow, you get one long take, not multishot.

3. Specify camera per shot. Each shot should have its own camera description (close-up, wide, tracking, static). Shots defined only by action tend to merge into continuous footage.

4. Include audio per shot. LTX-2.5 generates synchronized audio with video. If dialogue or sound is part of the scene, describe it per shot — who speaks, what is heard, when. Voice persistence across cuts is one of the advertised multishot properties, and it only has something to hold onto if you specified the voice.

5. Keep it under 200 words and chronological. The official guidance caps prompts at 200 words and wants a single flowing chronological paragraph. A 3–4 shot multishot prompt with full per-shot detail will get close to that ceiling — use the prompt enhancer for expansion rather than writing a novel.

A minimal working example, using the style from the official Quick Start:

A medium close-up of a woman with short dark hair, wearing a red jacket, standing in a sunlit cafe. She looks up from her coffee and smiles. Cut to a close-up of the same woman from the other side, the same red jacket and the same afternoon light behind her, as she says "it's ready." The camera then pulls back to a wide shot of the same cafe, the woman walking toward the door.

Note what this prompt does: identity ("the same woman"), wardrobe ("the same red jacket"), environment ("the same cafe"), and lighting ("the same afternoon light") are explicitly carried into each shot, and the cuts are marked ("Cut to…", "The camera then pulls back…").

Multishot in ComfyUI and ltx-pipelines

• ComfyUI: multishot is a model capability, not a separate workflow. The three official LTX-2.5 workflows (T2V, I2V, FLF2V) all generate with the model; prompting for multishot is what produces multi-shot output. The workflow files and model placement are covered in the ComfyUI LTX-2.5 tutorial.
• ltx-pipelines: generation happens through the same pipelines (DistilledPipeline etc.); there is no separate multishot flag. Omit --num-frames to let the duration head pick a length appropriate to your multi-scene prompt (frames must satisfy frames % 8 == 1 if you set them yourself).

FAQ

Is LTX 2.5 multishot available in every workflow? Yes — it is a capability of the model itself, so T2V, I2V, and FLF2V all inherit it. Whether a given prompt produces usable cuts is a prompting matter.

Can I combine multishot with image-to-video? Yes. I2V with a first-frame image, prompted as a multi-shot sequence, animates the scene with cuts. ComfyUI's I2V workflow is the same model.

Does multishot keep the same voice across shots? The model card lists voice among the properties held across cuts, alongside identity, environment, lighting, and style. Specify the voice in your prompt for it to have anything to hold onto.

Do I need the duration head for multishot? No — you can set --num-frames yourself. But multi-scene prompts imply longer clips, and the optional duration head is designed to predict length from the prompt, which is convenient for multi-shot prompts.

Bottom line

Multishot is LTX-2.5's structural upgrade: connected shots with identity, environment, lighting, voice, and style held across cuts — generated in one pass instead of stitched from separate runs. Prompt it as a marked shot list, carry identity and environment explicitly across shots, and keep everything within the 200-word chronological style the LTX-2 family is designed around. Test it cheaply in your browser at ltx23.app before you migrate your local setup.

Sources

Verified August 13, 2026 against primary documentation:

1. Lightricks/LTX-2.5 — Hugging Face model card — native multishot definition (identity, environment, lighting, voice, style across cuts), Gemma 4 12B encoder, prompt enhancer, duration predictor, limitations
2. LTX-2.5 official model page (ltx.io) — multishot capability description, prompt adherence claims
3. Lightricks/LTX-2 — GitHub README — prompting structure (under 200 words, chronological, shot-list style), enhance_prompt parameter, duration head usage
4. ComfyUI official LTX-2.5 tutorial — three native workflows, prompting tips (describe the whole scene, per-shot audio)
5. LTX-2 prompting guide (ltx.io blog) — official prompting methodology for the LTX-2 family

Multishot persistence is a vendor-described capability; results vary by prompt and hardware. Verify on your own generations.