LTX 2.5 Prompting: Why a Short Prompt Turns Into a Selfie Video

The apple that became a woman

Our first LTX 2.5 test was deliberately simple: "A single red apple resting on a white table, bright studio light, static shot." Same seed, same settings, two runs. Both came back as a woman.

A second test confirmed the pattern. "A wooden chair in an empty room, static shot" produced a woman opening a drink can.

Then we described the same apple in seven sentences — the waxy surface, the direction of the light, the shadow it casts, a droplet on the skin, the camera, the room tone. Nothing else changed. The result was a correct apple with speckled skin and a soft shadow.

The lesson: LTX 2.5 fills whatever you leave undescribed with its strongest prior, and that prior is a person on camera. A sparse prompt is not a minimal request. It is an invitation.

Write one flowing paragraph, in this order

LTX 2.5 wants prose, not a schema. The structure that bound best in our tests was a single paragraph of four to eight sentences, in this order:

1. Shot and framing — what the camera sees.
2. Scene and lighting — the space and where the light comes from.
3. Action, in chronological order — what happens, step by step.
4. Subject detail — textures, materials, surfaces.
5. Camera movement — or the lack of it, described concretely.
6. Audio — the soundscape you want to hear.

No JSON, no named fields, no bracketed audio tags. This matches the vendor's own guidance to write "a single flowing paragraph" of roughly four to eight descriptive sentences.

Abstract direction did no visible work. "Dramatic", "cinematic" and "epic" changed nothing. Concrete nouns and verbs did: motion, light direction, texture.

Settings: cfg 1.0, 8 steps — and why not 6

Run the distilled model at cfg 1.0 with the fixed 8-step schedule. Our test clips were 704×1280, 97 frames (about four seconds).

Six steps is a trap. Our first run at 6 steps gave the unrelated selfie; the same short prompt bound correctly at 8.

Raising cfg can force a thin prompt to bind, but it is the expensive way to fix the wrong problem. At cfg 3.0 a step took 91.3 seconds against 18.6 at cfg 1.0, with the card spilling into shared memory. Write a better prompt instead.

The negative prompt does nothing here

We carried a negative prompt through every run at cfg 1.0 and it never changed a result. In distilled mode at cfg 1.0 the negative prompt is ignored.

So write constraints as positive states. Instead of "no deformed hands", describe the hands you want: "her hands rest open and relaxed at her sides, fingers clearly separated."

Audio and dialogue

LTX 2.5 generates sound from the soundscape you describe, and the level follows the description closely. In our measurements a spoken line with gulls, wind and water came out at a mean of −18.6 dB; rain on metal with thunder at −33 dB; airport ambience at −40 dB. "Quiet room tone, soft rustle of fabric" measured −91 dB — effectively silent.

That last number matters: a silent audio track looks identical to a good one in a file inspector. Check loudness, not the presence of an audio stream.

Dialogue worked when the line was quoted, the speaker described and the clip kept to one short line per five seconds — for example: He looks straight into the lens, takes a breath and says clearly in a low weathered voice, "The tide turns at six."

Camera, objects and image-to-video

A locked camera depends on the subject. On a person, "locked-off tripod, does not move, zoom or tilt" still drifted: the framing tightened. On an object, the same instruction held for the whole clip.

Objects are where LTX 2.5 earned its place in our pipeline. On a chrome balloon plate where another model failed five of five prompts — abandoning the object after two seconds and inventing a person — LTX 2.5 bound the plate on the first attempt and held the framing.

Image-to-video honours the resolution you request: a 3:4 plate requested at 704×1280 came out at 704×1280. It costs more than text-to-video, though — 48 to 57 seconds per step against 18.6.

Clip length is a cost cliff

On a 16 GB card, frame count is not linear. At 97 frames a run took 7 minutes 42 seconds. At 125 frames — 29% more — it took 22 minutes 51 seconds, three times as long, and the decode stage hung for over 20 minutes without writing a file. We treat anything over 97 frames as a coin flip.

The model also adjusts frame counts: 97 requested came back as 105 in one keyframe test. Measure the output file.

Multi-shot prompts with named hard cuts ("Hard cut to shot two…") produced distinct angles, but objects did not stay consistent across cuts — a handleless cup became a handled espresso cup in the next shot. For a hero product, render each shot separately and cut them together in the edit.