# LTX-2 2.5 22B Distilled NVFP4

Profile `ltx25`, native ID `ltx2_25_22B_distilled_nvfp4`, is the default for every
currently supported video scenario. Other models are only candidates for future
unsupported scenarios. Do not switch models for OOM, slow execution or poor quality.

## Build a shot, then a prompt

Write concrete descriptions of the subject, place, visible action, camera, light
and style. Keep a character's appearance, clothes and surroundings consistent.
Describe events in the order they should occur. A short shot benefits from one
clear action rather than several scene changes. Use enough time for each action.

Use English visual instructions as a practical starting point. For dialogue put
the exact words in double quotes, identify the speaker and say what their mouth
and body do. Keep dialogue short enough for the selected duration. Russian output
needs listening/visual review; English wording is not a guarantee of good speech.
For controlled narration, generate speech separately, listen to it, then use
`condition` or `attach` according to the intended audio role.

For first-image shots, describe how the visible scene moves from that image.
Do not contradict its identity, composition or wardrobe. Prefer modest movement
for an initial take. If subtitles/text are not wanted, explicitly say so.

## Prompt Relay: one global prompt, timed local actions

Everything before the first valid bracket range is the global description for
the entire clip. Each range introduces the action for that part of the clip.
Ranges bias generation over time; they are not exact editing/keyframe commands.

| Range | Meaning |
| --- | --- |
| `[0%:25%]` | First quarter of the clip |
| `[25%:]` | From one quarter until the end |
| `[1:5]` | Frames 1 through 5, **not seconds** |
| `[0s:4s]` or `[0sec:4sec]` | Seconds |
| `[0:05:0:10]` | Timecode range: 5 to 10 seconds |

Use the same unit at both ends. Prefer percentages for the first attempts: native
frame rounding makes requested seconds differ slightly from final duration.
Keep ranges chronological and within the clip. Cover the full clip for continuous
action; compatible motion around boundaries avoids contradictory transitions.
Malformed markers may be treated as ordinary text by WanGP. Check them before
submitting. Use 2–3 purposeful beats for a short 3–6 second clip; more timed lines
do not automatically improve quality.

### Complete six-second example

```text
A woman in a yellow raincoat stands under a bus-stop roof on a quiet wet street. Her appearance and position remain consistent. Medium shot, slow gentle push-in, soft evening light, realistic motion. No captions or new visible text.
[0%:35%] She watches raindrops hit a small puddle and tilts her head toward a distant sound.
[35%:70%] She turns toward the camera, smiles, and says "Right on time." Her mouth moves naturally with the brief line.
[70%:] She looks toward the approaching bus, still smiling, as the camera settles.
```

Use `audio_mode: generate` if this line should be generated by LTX. In `none`
mode the delivered clip is silent. With supplied audio, make the timed movement
and quoted words compatible with that recording; the prompt cannot rewrite it.
If the clip must have no speech, remove the spoken line rather than relying only
on stripping the audio at export.

### Newlines: this API differs from the WebUI controls

Media Bridge forces WanGP `multi_prompts_gen_type: FG` — **all lines belong to one
prompt**. Pass the whole relay string as `prompt` in one JSON request. Newlines
and empty lines do not create extra clips, windows or jobs through this API.
The WebUI's paragraph/sliding-window recommendation applies when using those UI
modes directly; no line-processing flag is exposed by Media Bridge.

Native sliding-window/continuation semantics count frame 1 / 0% from the first
retained non-overlap frame of each window. This API currently exposes neither
continuation nor multi-window control. Do not infer those features from relay syntax.

## Profile settings

Current defaults: 8 inference steps, 2 guidance phases, NAG scale 2, tau 3.5,
alpha 0.5. Start with these defaults. Allowed request `tuning` fields are `steps`
(integer 1–50), `guidance` (0–10), `nag_scale` (0–5); the live schema is authoritative.
More steps are not automatically better for a distilled model. Compare one change
at a time with the same seed, dimensions, prompt and input assets. Reusing a seed
across different resolutions does not mean identical motion/composition.

Check the manifest's resolved model, effective settings and actual duration when
diagnosing a result. Do not present suggested prompt tactics as measured quality gains.

Basis: installed WanGP prompt relay parser and LTX handler, checked 2026-09-15;
[upstream prompt guide](https://github.com/deepbeepmeep/Wan2GP/blob/main/models/ltx2/prompt_enhancer.py).