LTX 2.5 Prompt Guide: Multishot, Audio, and the Gemma 4 Encoder

Every LTX prompt guide you will find for the next few weeks is a rewrite of the 2.3 guide with "2.5" swapped in. That misses what actually changed. LTX-2.5 swapped the text encoder for a custom Gemma 4 12B, added a prompt enhancer that rewrites your prompt before diffusion, and introduced multishot — connected scenes in one generation. Prompting is no longer only about what you write; part of the pipeline now writes for you.

The useful guide, then, has three parts: what the encoder expects from your prompt, what the enhancer does to it, and how to structure a multishot prompt with audio. Sourced from the LTX-2 README, the LTX-2.5 model card, the official prompting guide, and the ComfyUI LTX-2.5 tutorial. Checked August 13, 2026.

1. The base structure the LTX family expects

The official guidance (unchanged from LTX-2 and still the documented style for 2.5) is a detailed, chronological description of actions — one flowing paragraph, under 200 words, literal and precise, written like a cinematographer describing a shot list.

The official README's build order:

• Start with the main action in a single sentence
• Add specific movements and gestures
• Describe character/object appearances precisely
• Include background and environment details
• Specify camera angles and movements
• Describe lighting and colors
• Note any changes or sudden events

The official example in the README is a single long paragraph covering appearance, expression, sound, speech content, camera, depth of field, and background — one take, fully specified. That style remains the foundation for 2.5; what changes is how far you can push complexity.

2. What the Gemma 4 12B encoder changes

The model card describes the custom Gemma 4 12B encoder as holding complex prompts together — multiple characters, camera moves, lighting, actions — instead of dropping details across a longer sequence.

Practical consequences for your prompts:

• Longer, denser prompts survive. The old encoder dropped details in complex prompts; the Gemma 4 12B is explicitly designed to retain multiple subjects, actions, lighting, and camera direction. You can now write genuinely multi-part prompts without expecting the tail to get truncated.
• The encoder is version-checked, not optional. The README states loading checks the encoder's version against the checkpoint's (gemma4-12b-ltx-v1), and Google's stock Gemma 4 is not a substitute. If your pipeline uses a stock encoder, it fails the check — no silent degradation.
• There are two encoder files in ComfyUI. The official workflows load the LTX-tuned gemma4-12b-with-proj-ltx-2.5-comfy-int8-convrot and a second gemma4_e2b_it_bf16 from Comfy-Org. Missing either one breaks the workflow (see our ComfyUI guide).

What does not change: the 200-word cap is still documented, and prompt-following is still "heavily influenced by prompting style" per the model card's own limitations section. A better encoder raises the ceiling; it does not remove the need for structure.

3. The prompt enhancer: when to use it

The model card lists a prompt enhancer — expands a short prompt into richer cinematic instructions at minimal extra compute. The README exposes it as an enhance_prompt parameter on the pipelines.

The workflow implication: you can write a simple scene idea ("a knight walks through a foggy forest") and let the enhancer expand it into full cinematic instructions. The ComfyUI tutorial confirms this is the intended flow: "Write a simple scene idea and let the enhancer handle the rest."

The caveat is control. If the enhancer's cinematic expansion is not your taste — or you need a specific camera move, exact wardrobe, or precise dialogue — write the full prompt yourself and skip enhancement. The enhancer is a floor-raiser for short prompts, not a substitute for specifying what matters to you. Test both on the same idea once; the difference is instructive about what the model considers "cinematic."

4. Multishot prompting: the shot list

Multishot is the structural new capability — multiple connected shots in a single pass, holding character identity, environment, lighting, voice, and style across cuts. It has its own prompting requirements, covered in detail in our multishot guide; the essentials:

1. Introduce identity and environment once, then reference them ("the same man", "the same cafe")
2. Mark the cuts explicitly ("Cut to a close-up of…", "The camera then pulls back…")
3. Give each shot its own camera — shots defined only by action merge into one take
4. Describe audio per shot — dialogue, sound, who speaks when; voice persistence across cuts is a multishot property, and it needs the voice described
5. Stay under 200 words — with the enhancer available, write the spine and let it expand, rather than exceeding the documented cap

5. Audio prompting

LTX-2.5 generates synchronized audio with video, so sound belongs inside the prompt, not in a separate track. The ComfyUI tutorial's phrasing is direct: "Describe sounds and dialogue needed for the scene."

Audio prompting rules that work with the official style:

• Describe sounds as part of the chronological flow: "as she opens the door, a faint bell chimes"
• For dialogue, include the speaker and the tone: "he says in a low voice, 'we're too late'"
• For I2V, audio is part of "what happens next": motion, camera, and sounds that follow from the image
• For FLF2V, audio is part of the transition between the two frames

A complete example, structured for 2.5

Here is a multishot prompt that uses all three mechanisms — dense encoder input, marked cuts, and per-shot audio:

A close-up of a woman with short dark hair in a red jacket, standing in a sunlit cafe near the window. She looks down at her coffee and smiles faintly as rain starts against the glass, a soft patter audible. Cut to a medium shot of the same woman, the same red jacket and the same afternoon light, as she says quietly, "it's ready now." The camera slowly dollies left past her to a wide shot of the same cafe, the rain heavier, and a man in a grey coat entering through the door, the bell chiming once.

Note what carries across: identity ("the same woman"), wardrobe ("the same red jacket"), environment ("the same cafe"), and light ("the same afternoon light"), with explicit cuts and per-shot audio. That is the structure the model's documented multishot property is designed to preserve.

Test your prompting live. Write your prompt, generate with LTX 2.5 — audio included — at ltx23.app, and iterate on structure before you tune anything locally.

FAQ

Do I still need to keep prompts under 200 words? Yes — the 200-word chronological paragraph style is the documented structure. The Gemma 4 12B encoder retains complex details better, but the cap stands, and the enhancer is the sanctioned way to expand.

Should I use the prompt enhancer or write full prompts? Short scene ideas → let the enhancer expand. Precise creative control (exact camera, wardrobe, dialogue) → write the full prompt yourself. The enhancer raises the floor; it does not remove your need for specificity.

How do I prompt for audio in LTX 2.5? Describe sounds and dialogue inside the chronological prompt — as part of the action, not as a separate instruction list. For multishot, describe audio per shot.

How is LTX 2.5 prompting different from LTX 2.3? The base structure is the same (chronological, under 200 words, shot-list style). What changed: the Gemma 4 12B encoder retains denser prompts, the prompt enhancer expands short prompts automatically, and multishot prompts need explicit cuts and per-shot details.

Does the stock Gemma 4 work as the text encoder? No. The LTX-tuned encoder (gemma4-12b-ltx-v1 version) is required and version-checked.

Bottom line

LTX 2.5 prompting keeps the family structure — one flowing chronological paragraph under 200 words — and adds three mechanisms: a Gemma 4 12B encoder that retains denser prompts, a prompt enhancer for short ideas, and multishot prompting with marked cuts and per-shot audio. Write the spine yourself, let the enhancer expand when you don't need precision, and describe sound as part of the action. Iterate your prompts in the browser at ltx23.app before tuning locally.

Sources

Verified August 13, 2026 against primary documentation:

1. Lightricks/LTX-2 — GitHub README — prompting structure (under 200 words, chronological, shot-list), enhance_prompt parameter, Gemma 4 12B version check
2. Lightricks/LTX-2.5 — Hugging Face model card — Gemma 4 12B encoder, prompt enhancer, multishot, prompting limitations
3. Official LTX-2 prompting guide (ltx.io blog) — official prompting methodology for the LTX-2 family
4. ComfyUI official LTX-2.5 tutorial — per-workflow prompting tips (whole scene, what happens next, transition), prompt enhancer note
5. LTX-2.5 official model page (ltx.io) — prompt adherence claims, multishot description

Prompt-following varies with prompting style per the model's own limitations section; results differ by hardware and prompt. Verify on your own generations.