P-Video-2-Pro

Introduction

Every P-Video-2-Pro clip comes back with a soundtrack generated in the same pass as the picture. Voices, effects, music and room tone are all written from the prompt, so a line lands on the mouth that says it and a door sounds on the frame it closes.

Turn the sound on:

Two people, one quoted line each, delivered word for word and in turn with the mouths matching. This guide covers what happens to sound you do not write, how to direct the mix, how to time effects to the action, how spoken lines behave, how to stage two speakers, and how to get a sung vocal.

Sound on every clip

There is no audio switch on this model. A request carrying settings.audio is rejected as an unsupported parameter, and every clip returns with a track. What you control is what that track contains, and the prompt is the only place to say it.

The same product shot with and without an Audio: clause, on one seed:

Without the clause the model writes a track of its own, and here that is a music bed playing at full level under a product shot nobody asked to score. With it, the clip is near silence apart from one soft click. The picture changes as well, since the audio clause is part of the same prompt.

For a clip that must ship silent, ask for near silence in the prompt, then remove the audio stream in your own pipeline. Voiceover and licensed music go on in the edit, over a picture that was never asked to carry them.

Directing the mix

Everything after Audio: describes the track. The same pizza shot with only that clause changed:

The picture brief is identical in all three, and the tracks are not. Each clause names the layer that leads, the crunch, the groove or the kitchen, and writes the crunch in as well. How present the crunch stays depends on what it shares the track with and how loudly the clause asks for it.

Write each layer with its level attached:

• Effects are tied to something on screen and described as close or in the foreground when they need to be heard.
• Music is named by genre, instrument and tempo, and by its place in the mix: leading, or under something else.
• Ambience is the place and the things making noise in it, not the word "ambience" on its own.
• Speech goes in quotation marks and has its own sections below.

Timing effects to the action

A sound listed on its own has no moment attached to it. Tie each effect to the event that makes it, "each time a glove lands on a pad", and the model has a frame to put it on. The same pad session, both ways:

In the tied version a smack lands with every punch, from the first combination to the last, because the clause gave every sound its trigger. The listed version plays punch sounds in bursts, with about three seconds in the middle where he keeps punching and nothing is heard.

The pattern is "the sound, then the event that makes it": a click as the hinge moves, a smack each time a glove lands. It costs a few words per effect.

Spoken lines

Put every line in quotation marks and attribute it to someone in the shot. A quoted line is spoken as written. A described line ("she explains the tip") gets the topic and the model's own wording, which is no use for a price or a product name.

Length is the other half. The same coaching tip at two lengths, on one seed:

The short line sits inside eight seconds with room for a beat before and after it, and the mouth follows every word. The long one asks for forty words in five seconds, and the delivery has to race to fit before the clip ends.

Count the words before you set duration. Natural speech runs at roughly two to three words a second, so leave a second at each end on top of that. When a script runs longer, split it at sentence breaks across several clips and cut them together.

Two more details shape the performance. Name the delivery ("clear upbeat", "dry") the way you would brief a presenter. And write what happens after the line, "then she smiles and nods once", so the clip does not end on a held mouth.

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-pro',
  positivePrompt: 'A fitness coach in her twenties in a black training top stands in a bright gym beside a cable machine, looks into the lens and says, "Keep your elbows tucked, and slow down on the way back up." Then she smiles and nods once. Vertical creator framing, chest-up, handheld with very slight movement, bright even gym lighting. Photoreal fitness creator cinematography, no text, no logos. Audio: her clear upbeat speaking voice, close to the microphone, quiet gym ambience underneath, no music.',
  width: 768,
  height: 1344,
  duration: 8,
  seed: 4402918,
  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-pro",
            "positivePrompt": "A fitness coach in her twenties in a black training top stands in a bright gym beside a cable machine, looks into the lens and says, \"Keep your elbows tucked, and slow down on the way back up.\" Then she smiles and nods once. Vertical creator framing, chest-up, handheld with very slight movement, bright even gym lighting. Photoreal fitness creator cinematography, no text, no logos. Audio: her clear upbeat speaking voice, close to the microphone, quiet gym ambience underneath, no music.",
            "width": 768,
            "height": 1344,
            "duration": 8,
            "seed": 4402918,
            "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": "4b9e1d73-6a25-4c08-8f31-d2a7c5e9b160",
      "model": "prunaai:p-video@2-pro",
      "positivePrompt": "A fitness coach in her twenties in a black training top stands in a bright gym beside a cable machine, looks into the lens and says, \"Keep your elbows tucked, and slow down on the way back up.\" Then she smiles and nods once. Vertical creator framing, chest-up, handheld with very slight movement, bright even gym lighting. Photoreal fitness creator cinematography, no text, no logos. Audio: her clear upbeat speaking voice, close to the microphone, quiet gym ambience underneath, no music.",
      "width": 768,
      "height": 1344,
      "duration": 8,
      "seed": 4402918,
      "deliveryMethod": "async"
    }
  ]'

runware run prunaai:p-video@2-pro \
  positivePrompt="A fitness coach in her twenties in a black training top stands in a bright gym beside a cable machine, looks into the lens and says, \"Keep your elbows tucked, and slow down on the way back up.\" Then she smiles and nods once. Vertical creator framing, chest-up, handheld with very slight movement, bright even gym lighting. Photoreal fitness creator cinematography, no text, no logos. Audio: her clear upbeat speaking voice, close to the microphone, quiet gym ambience underneath, no music." \
  width=768 \
  height=1344 \
  duration=8 \
  seed=4402918 \
  deliveryMethod=async

{
  "taskType": "videoInference",
  "taskUUID": "4b9e1d73-6a25-4c08-8f31-d2a7c5e9b160",
  "model": "prunaai:p-video@2-pro",
  "positivePrompt": "A fitness coach in her twenties in a black training top stands in a bright gym beside a cable machine, looks into the lens and says, \"Keep your elbows tucked, and slow down on the way back up.\" Then she smiles and nods once. Vertical creator framing, chest-up, handheld with very slight movement, bright even gym lighting. Photoreal fitness creator cinematography, no text, no logos. Audio: her clear upbeat speaking voice, close to the microphone, quiet gym ambience underneath, no music.",
  "width": 768,
  "height": 1344,
  "duration": 8,
  "seed": 4402918,
  "deliveryMethod": "async"
}

[
  {
    "taskType": "videoInference",
    "taskUUID": "4b9e1d73-6a25-4c08-8f31-d2a7c5e9b160",
    "videoUUID": "e61c8a24-9f53-4b17-a2d8-5c3e7b9f0a46",
    "videoURL": "https://vm.runware.ai/video/os/a14d18/ws/2/vi/e61c8a24-9f53-4b17-a2d8-5c3e7b9f0a46.mp4"
  }
]

Two speakers

Two people talking need two anchors: a position in frame and something visible that tells them apart, such as "the patient on the left, in a denim jacket". Then give each one a single quoted line and say the voices take turns. The same exchange staged two ways:

Both keep each line on the right face, and both deliver the lines word for word. The two-shot holds both people in one frame, so the listener's reaction is part of the shot, which suits a sketch or a testimonial exchange. The numbered version cuts to an over-the-shoulder close-up for each line, the coverage a drama scene uses, and gives the speaker the frame while they talk.

Keep it to one line per speaker per clip. A longer conversation is several clips cut together, each with its own pause, which is also where the edit points want to be. Writing prompts covers numbering shots in more detail.

Sung vocals

The model writes melody as well as speech. Quote the words, say they are sung rather than spoken, and name the instrument playing with them:

The guitar and the voice come from the same generation, so the strumming hand and the singing mouth share one timing. The melody is the model's own. A clip that has to match an existing song needs that song as an input, which this model does not take.

Tips

1. Write the audio clause every time. There is no off switch, so a clip without one gets whatever track the model picks.

2. Strip the track for silent deliverables. Ask for near silence in the prompt, then remove the audio stream before the voiceover goes on.

3. Name the layer that leads. Say which of effects, music, ambience or speech sits on top, and describe the rest relative to it.

4. Tie every effect to its event. "A smack each time a glove lands" puts the sound on the frame. "Punches" does not.

5. Quote every line that has to be exact. A described line gets the model's wording.

6. Count the words against the duration. Roughly two to three words a second, plus a second at each end.

7. Anchor each speaker. Position in frame plus something visible, and one quoted line each.

8. Say sung, not said. Quote the lyric, name the instrument, and the model writes the melody around it.