P-Video-2

Introduction

Most video models hand you a silent file and leave the soundtrack to your edit. P-Video-2 generates the audio with the picture, in the same call, from the same prompt, which means the sound is synchronised because it was never separate. A door closes on the frame the door closes. A line lands with the mouth that says it.

That last part is the one worth testing first. Turn the sound on:

Notice the shape of the delivery rather than the words. She pauses, delivers the line, then rests, and the mouth follows the whole way. This guide covers the flag that controls it, how to direct each layer of the mix, how spoken lines behave, how to put two people in one scene, and when to ask for silence instead.

The flag

settings.audio is a boolean and it defaults to true, so every call already returns a video with a soundtrack whether you planned one or not.

import { createClient } from '@runware/sdk'

const client = await createClient({ apiKey: process.env.RUNWARE_API_KEY })
await client.connect()

const [result] = await client.run({
  model: 'prunaai:p-video@2',
  positivePrompt: 'A close-up of a cream stand mixer on a pale marble kitchen counter, its dough hook turning steadily in a steel bowl of pale dough, morning daylight from a window to the right. Photoreal product demonstration cinematography, locked-off camera, no text, no people. Audio: a powerful kitchen mixer running hard, its motor droning loudly and the dough thudding rhythmically against the metal bowl, recorded close so both are unmistakable. No music, no voice.',
  width: 1280,
  height: 704,
  duration: 6,
  settings: {
    audio: true
  },
  seed: 3390471,
  deliveryMethod: 'async'
})

import asyncio
import os

from runware import Runware


async def main():
    async with Runware(api_key=os.environ["RUNWARE_API_KEY"]) as client:
        results = await client.run({
            "model": "prunaai:p-video@2",
            "positivePrompt": "A close-up of a cream stand mixer on a pale marble kitchen counter, its dough hook turning steadily in a steel bowl of pale dough, morning daylight from a window to the right. Photoreal product demonstration cinematography, locked-off camera, no text, no people. Audio: a powerful kitchen mixer running hard, its motor droning loudly and the dough thudding rhythmically against the metal bowl, recorded close so both are unmistakable. No music, no voice.",
            "width": 1280,
            "height": 704,
            "duration": 6,
            "settings": {
                "audio": True
            },
            "seed": 3390471,
            "deliveryMethod": "async"
        })


asyncio.run(main())

curl https://api.runware.ai/v1 \
  -H "Authorization: Bearer $RUNWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[
    {
      "taskType": "videoInference",
      "taskUUID": "e1a7c395-4d82-4b06-9f37-2c8e5a1d7b40",
      "model": "prunaai:p-video@2",
      "positivePrompt": "A close-up of a cream stand mixer on a pale marble kitchen counter, its dough hook turning steadily in a steel bowl of pale dough, morning daylight from a window to the right. Photoreal product demonstration cinematography, locked-off camera, no text, no people. Audio: a powerful kitchen mixer running hard, its motor droning loudly and the dough thudding rhythmically against the metal bowl, recorded close so both are unmistakable. No music, no voice.",
      "width": 1280,
      "height": 704,
      "duration": 6,
      "settings": {
        "audio": true
      },
      "seed": 3390471,
      "deliveryMethod": "async"
    }
  ]'

runware run prunaai:p-video@2 \
  positivePrompt="A close-up of a cream stand mixer on a pale marble kitchen counter, its dough hook turning steadily in a steel bowl of pale dough, morning daylight from a window to the right. Photoreal product demonstration cinematography, locked-off camera, no text, no people. Audio: a powerful kitchen mixer running hard, its motor droning loudly and the dough thudding rhythmically against the metal bowl, recorded close so both are unmistakable. No music, no voice." \
  width=1280 \
  height=704 \
  duration=6 \
  settings.audio=true \
  seed=3390471 \
  deliveryMethod=async

{
  "taskType": "videoInference",
  "taskUUID": "e1a7c395-4d82-4b06-9f37-2c8e5a1d7b40",
  "model": "prunaai:p-video@2",
  "positivePrompt": "A close-up of a cream stand mixer on a pale marble kitchen counter, its dough hook turning steadily in a steel bowl of pale dough, morning daylight from a window to the right. Photoreal product demonstration cinematography, locked-off camera, no text, no people. Audio: a powerful kitchen mixer running hard, its motor droning loudly and the dough thudding rhythmically against the metal bowl, recorded close so both are unmistakable. No music, no voice.",
  "width": 1280,
  "height": 704,
  "duration": 6,
  "settings": {
    "audio": true
  },
  "seed": 3390471,
  "deliveryMethod": "async"
}

[
  {
    "taskType": "videoInference",
    "taskUUID": "e1a7c395-4d82-4b06-9f37-2c8e5a1d7b40",
    "videoUUID": "7c2f9a48-1e63-4d50-8b19-4a7d3c1f9e26",
    "videoURL": "https://vm.runware.ai/video/os/a14d18/ws/2/vi/7c2f9a48-1e63-4d50-8b19-4a7d3c1f9e26.mp4"
  }
]

The consequence of that default is worth stating plainly. A prompt that says nothing about sound still gets sound, chosen by the model to suit the picture, which for a corporate explainer is usually a music bed nobody cleared. The Audio: clause is how you take that decision back.

Directing the mix

Everything after Audio: is read as a description of the track. The same product shot, on one seed, with only that clause changing:

Three different tracks from one picture: near-silence, a working machine, and a music bed.

What decides which one you get is how loudly you say it, and the model takes those words literally. Ask for "quiet kitchen room tone" and you get effectively nothing, close enough to digital silence to read as a muted file. The middle clip only has an audible motor because the clause says droning loudly and recorded close so both are unmistakable. An early draft of this guide asked for a "low steady hum" and a "soft rhythmic slap", and the result was indistinguishable from the silent version.

So write the level as deliberately as the source:

• Ambience is what you get by default and what you get when you ask for anything quiet. Useful when you want presence and nothing more, and a trap when you actually wanted to hear something.
• Effects need to be told they are loud, close, or in the foreground, or they sink under the noise floor.
• Music responds to genre, instrument and tempo, and to its position in the mix. "A light acoustic instrumental bed with a plucked guitar figure" gets you closer than "upbeat music".
• Speech is the strongest layer and has its own section.

The third clip makes the same point from the other side. Its clause asks for a music bed with the mixer motor low underneath it, and what comes back is the music alone. Asking for a layer quietly is close to asking for it to be dropped.

Position words like "under" and "over" are read, but they do not rescue a layer you have also described as quiet. Ask for both sounds at a level you would actually mix them at, and say which one leads.

Speech, and why the mouth matches

Native speech is the reason to reach for this model over a silent one, and it works best when you quote the line you want. The same setup, quoted and paraphrased:

Both men talk. Only one of them says what you wrote. The paraphrased version gets the subject matter and invents the sentence, which is fine for background chatter and useless for a claim that has to be legally accurate or a price that has to be right.

The delivery has a consistent shape worth designing around. The model pauses, delivers, then rests, so a spoken clip is a beat of silence, the line, and a beat of silence, and the mouth tracks the audio through all three. Two things follow from that.

Leave room for the beats. A duration that fits only the words gives you a clip that starts mid-syllable. Budget a second at each end and the delivery sits properly inside the clip.

Write the rest into the prompt. "Then he rests", "then she lowers the tube and smiles" gives the model something to do with the tail instead of holding a frozen face.

A line that runs long shows exactly that:

The content survives and the performance does not. Split a script across clips at its natural sentence breaks rather than fitting more words into one call, then cut them together. Each clip gets its own pause and rest, which is also where your edit points want to be.

Two voices in one scene

Two speakers in one frame is a genuine capability and the thing an avatar model cannot do. Give each person their own quoted line and say they take turns:

The woman on the left, in a green blazer, says, "So what actually changed
this quarter?" The woman on the right, in a white shirt, answers, "Support
tickets dropped by half."

Audio: two distinct natural speaking voices taking turns without
overlapping, quiet studio tone underneath, no music.

Anchoring each speaker to a position in frame and a garment is what keeps the voices attached to the right faces. "The woman on the left, in a green blazer" is doing the same job here that it does in a removal instruction: separating one subject from another that looks similar.

A two-hander is the shape to reach for, and it is the thing a single-presenter avatar model cannot give you at all. Give each speaker one quoted line and say they take turns, and the anchoring does the rest.

For a scene that needs a third voice, the reliable route is to shoot them as their own clip and cut, the same way you would split a long script. That keeps every line attached to the face you meant.

Turning the sound off

settings.audio: false returns the video with no audio track, and there are real reasons to ask for it.

{
  "settings": {
    "audio": false
  }
}

Social feeds autoplay muted, and a listing loop that will never be heard does not need a soundtrack. More importantly, a clip destined for a voiceover should arrive silent, because a generated music bed underneath your own narration is a mix problem you have to solve in the edit.

That prompt has no Audio: clause, which is the right pairing. With the flag off there is nothing for the clause to direct, and a sound description left in a silent request is just prompt length competing with the picture.

Production patterns

A training scene with a full mix. This is the shape an avatar model cannot deliver: an instruction that has to be exact, an environment that has to read as a real workplace, and a bed underneath it.

Four things at once: a voice, an underscore, warehouse ambience and the trolley. That is at the top of what one call reliably layers, so check this one rather than assuming it. The trolley is the layer most likely to land, because you can see it, and a forklift beeping somewhere off camera would have nothing to sync to.

A UGC ad with the line over a beat. Dialogue on top of music is the combination that sends this to P-Video-2 rather than an avatar, and it is two layers, which is well inside what the model will genuinely mix.

Two details are carrying that one. Naming the delivery ("dry deadpan") shapes the performance the way naming a light shapes an image, and three short phrases instead of one sentence gives the pacing its punctuation without asking the model to rush.

Tips

1. Always write the audio clause. The flag defaults on, so silence and music are both choices you make. "No music, no voice" is as much a direction as a genre is.

2. Quote every line that has to be exact. A described line gets you the topic and the model's own wording. Prices, claims and product names all need the quotes.

3. One breath per clip. Write what a person would say without pausing, then split the script and cut. Compressed pacing is where lip sync fails first.

4. Budget a second at each end. The delivery is pause, line, rest. A duration cut to the words alone clips the shape off both sides.

5. Write the tail. "Then he rests", "then she lowers the tube" stops the clip ending on a frozen face.

6. Keep effects visible and few. Two on-screen effects sync. Six effects, half of them off camera, land at random. Do the rest as foley in your edit.

7. Anchor every speaker. Position in frame plus a garment, and one quoted line each. That is what keeps a voice attached to the right face when two people share a shot.

8. Ask for silence when a voiceover is coming. settings.audio: false is cheaper than fighting a generated music bed underneath your own narration.

9. Two or three layers mix, six queue. Say explicitly that the sounds run at the same time, keep the count low, and put anything beyond that in your edit.