LTX 2.3 Prompt Guide: Write Prompts That Work

Most LTX 2.3 prompt advice is recycled Stable Diffusion advice: comma-separated tags, quality boosters, a pile of adjectives. It produces exactly what you would expect — a video that looks like a moving photograph.

LTX 2.3 wants something different, and Lightricks documents it plainly. The official guidance asks for detailed, chronological descriptions in a single flowing paragraph, written the way a cinematographer describes a shot — literal, precise, and within 200 words.

That last constraint surprises people. A 200-word ceiling, in a model that rewards detail, means every word has to earn its slot. This guide covers the official structure, what to put in each slot, how to prompt the audio track (LTX 2.3 generates synchronized audio, which changes prompting entirely), and where the built-in enhancer fits.

Sources: the LTX-2 official repository prompting section, Lightricks' prompting guide, and the pipeline documentation, checked July 2026.

The official structure

Lightricks publishes a seven-part order to build prompts in:

1. Start with the main action in a single sentence
2. Add specific details about movements and gestures
3. Describe character and object appearances precisely
4. Include background and environment details
5. Specify camera angles and movements
6. Describe lighting and colors
7. Note any changes or sudden events

Plus three framing rules: start directly with the action, keep descriptions literal and precise, and keep the whole thing in one flowing paragraph under 200 words.

Read that list again and notice what is missing: no style tags, no quality boosters, no artist names, no negative-prompt scaffolding. The structure is a shot description, not a tag soup.

Why "chronological" is the word that matters

The single biggest upgrade to most prompts is treating the prompt as a timeline, not a description of a frozen moment.

A static prompt gives you a static result:

A woman in a red coat standing in a train station, cinematic lighting, 4k, highly detailed

A chronological prompt gives the model something to animate:

A woman in a red wool coat walks briskly through a train station concourse, checking her watch mid-stride, then slows as she spots the departure board. Commuters blur past on both sides. The camera tracks her from the front at chest height, pulling back to keep her centered, while overhead fluorescents throw hard white light across the polished floor. As she stops, her expression shifts from urgency to relief.

Same subject, same scene. The second version specifies what happens first, what happens next, and what changes — which is what a video model has to decide anyway. Leaving it unspecified does not give you creative freedom; it gives the model the freedom to pick something boring.

"4k, highly detailed" bought nothing in the first version. That is the space where the change of expression fits.

Prompting the audio track

This is where LTX 2.3 diverges hardest from image-model habits, and where most third-party prompt guides simply have nothing to say.

LTX-2 is described by Lightricks as a DiT-based audio-video foundation model with synchronized audio and video generated together. So the audio is not an afterthought you add later — it is part of the same prompt.

The official quickstart prompt in the LTX-2 README demonstrates this directly. Inside a single shot description, it specifies:

• an incidental sound event — the text notes that "a quick sniff sound is heard"
• the voice quality and tone — the subject speaks with a deep male voice in a satisfied tone
• the actual dialogue lines, written out in quotes
• a pause, then a second sound event, then more dialogue

That structure is the template. When you want speech, write it like a script beat inside the shot description: who speaks, how their voice sounds, what they say verbatim, and what non-speech sounds happen around it.

Practical rules that follow from it:

• Put dialogue in quotation marks. It signals "these exact words are spoken."
• Describe the voice, not just the words. Age, register, tone, accent, pace — the model has no other way to choose.
• Name ambient sound explicitly. Rain on metal, a room's echo, distant traffic. Silence you did not ask for is a choice the model makes for you.
• Sequence the audio with the action. "After a brief pause" is doing real work in that official example — it places a sound in time.

For audio-only work, there is even a dedicated pipeline: T2AOneStagePipeline generates speech or audio from text with no video branch at all, useful for testing a voice LoRA before spending compute on frames.

Iterating on prompt phrasing? Prompt work is a numbers game — you need many quick attempts, not a few expensive ones. ltx23.app lets you run LTX 2.3 prompts in the browser, which is a faster loop for finding phrasing that lands than rerendering locally each time.

The 200-word budget, allocated

If 200 words is the ceiling, here is a division that works:

The value of a budget is that it makes you notice waste. Every "masterpiece, best quality, ultra-realistic" is roughly a camera move you did not get to specify.

Camera language that works

Vague camera direction is where a lot of otherwise-good prompts get flat results. Compare "the camera moves" against naming an actual move:

• Static — locked off, no movement
• Tracking / dolly — follows the subject through space
• Push in / pull back — camera moves toward or away from the subject
• Handheld — subtle instability, documentary feel
• Jib up / down — vertical crane movement
• Orbit — arcs around the subject

That vocabulary is not arbitrary — Lightricks ships camera-control LoRAs named after these moves (dolly in, dolly out, dolly left, dolly right, jib up, jib down, static). When a lab trains a dedicated adapter per move, that is a strong signal the model already recognizes the terms in text.

Also specify camera height and distance: "medium close-up at eye level" is a real instruction; "cinematic shot" is not.

Negative prompts and the enhancer

Two features worth knowing:

Prompt enhancement. LTX-2 pipelines support automatic prompt enhancement via an enhance_prompt parameter, exposed on the CLI as --enhance-prompt, which rewrites your prompt before generation. It is a reasonable safety net for short prompts, and a bad idea once you have a carefully budgeted 200-word description — it will rewrite the thing you tuned.

Negative prompts exist in the stack — the text-to-audio pipeline documents a --negative-prompt flag among its optional arguments — but they are not the centerpiece the way they are in image workflows. The official prompting guidance says nothing about building an elaborate negative prompt, and the honest reading is: fix the positive prompt first. Most "artifacts" people try to negative-prompt away are actually under-specification in the shot description.

A checklist before you hit generate

• Does it start with an action, not a noun phrase?
• Is it one paragraph, under 200 words?
• Does something change between the start and end?
• Is the camera move named, with height and distance?
• Is the lighting sourced (window, neon, overcast) rather than adjectival ("cinematic")?
• If anyone speaks, are the words in quotes and the voice described?
• Did you delete every quality booster?
• Is the resolution divisible by 32 and the frame count on the 8k+1 pattern? (Model constraints — they will bite regardless of prompt quality.)

FAQ

How long should an LTX 2.3 prompt be? Under 200 words, per Lightricks' own guidance, in a single paragraph.

Do quality tags like "4k, masterpiece" help? They are not part of the official guidance. The documented approach is literal, chronological shot description.

How do I get dialogue in the output? Write the lines in quotes inside the prompt, and describe the speaker's voice quality and tone — the official example prompt does exactly this.

Should I use the prompt enhancer? Useful for thin prompts, counterproductive for tuned ones. It rewrites what you wrote.

Is there an official prompt guide? Yes — the prompting section of the LTX-2 README plus Lightricks' published prompting guide at ltx.io.

Bottom line

LTX 2.3 prompting is closer to writing a shot list than to writing tags: one paragraph, chronological, literal, under 200 words, with the camera and audio specified as deliberately as the subject.

The fastest way to internalize it is repetition — write a shot, watch what the model did with the parts you left vague, tighten those. Run that loop in the browser at ltx23.app and you will develop the instinct in an afternoon rather than a week.

Sources

Verified July 2026 against primary documentation:

1. Lightricks/LTX-2 — official GitHub repository — the seven-part prompt structure, 200-word limit, quickstart example prompt, prompt enhancement
2. Lightricks official prompting guide — the guide the README points to
3. LTX-2 available pipelines documentation — T2A pipeline, negative prompt flag, audio conditioning behavior
4. LTX-2 installation & CLI flags — --enhance-prompt and shared flags
5. Lightricks/LTX-2.3 — Hugging Face model card — resolution and frame-count constraints, audio-visual generation
6. LTX-2 camera control LoRAs (dolly in) — named camera-move adapters
7. LTX-2 camera control LoRAs (jib up) — named camera-move adapters
8. ComfyUI official LTX-2.3 tutorial — workflow context for prompt inputs
9. LTX-2 optimization documentation — step counts affecting iteration speed
10. LTX-2 technical paper (arXiv 2601.03233) — joint audio-video architecture
11. ltx.io — official LTX site — product and model documentation hub

Example prompts in this article are written for illustration; the structural rules and the audio-prompting pattern are taken from Lightricks' published documentation.