PixVerse V6

Introduction

Multi-shot reels are the awkward middle ground of generated video. A single text-to-video call produces one continuous take, which is fine for an atmospheric clip but wrong for anything that needs to cut between angles, locations, or moments. The usual workaround is the editor's bench: generate three or four shots separately, line them up on a timeline, match the color grade by hand, hope the subject reads as the same person or object across cuts.

PixVerse V6 collapses that workflow into a single call. You hand it one or two anchor frames, write a prompt that names the beats of the story, flip settings.multiClip on, and the model returns a multi-shot reel with the cuts in place and the subject carried through. The anchor frames lock the visual identity at the moments you care about. The prompt drives what happens between them.

The reel above was generated in one API call from the two anchor frames shown below. The model worked out the camera angles for the middle shots, kept the color grade consistent across cuts, and matched the bracketing frames at the open and close.

A misty Mediterranean hillside vineyard at sunrise, rows of trellised grapevines stretching across a gentle slope into soft pastel light, fine dewdrops catching the early sun on the leaves, low fog hanging between the rows, a single woven wicker basket on the dirt path

A single tall stemmed glass of deep ruby red wine resting on a weathered wooden table at golden hour, rolling Mediterranean vineyard slopes blurred softly into the warm sunset background, soft lens flare on the rim of the glass

This guide covers the three input shapes V6 supports, the beat-structured prompt pattern that drives clean cuts, how to use a pair of anchors for a smooth transition instead of a reel, when the thinking parameter is worth flipping on, and how to think about the audio surcharge.

Request shape

A V6 multi-shot request takes a prompt, one or two anchor frames in inputs.frameImages, and a small set of flags in settings:

import { createClient } from '@runware/sdk'

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

const [result] = await client.run({
  model: 'pixverse:1@8',
  positivePrompt: 'A cinematic multi-shot sequence... Start with... Cut to... Shift to... End with...',
  inputs: {
    frameImages: [
      {
        image: 'https://example.com/first.jpg',
        frame: 'first'
      },
      {
        image: 'https://example.com/last.jpg',
        frame: 'last'
      }
    ]
  },
  duration: 10,
  resolution: '1080p',
  settings: {
    multiClip: true,
    audio: true,
    thinking: 'enabled'
  }
})

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": "pixverse:1@8",
            "positivePrompt": "A cinematic multi-shot sequence... Start with... Cut to... Shift to... End with...",
            "inputs": {
                "frameImages": [
                    {
                        "image": "https://example.com/first.jpg",
                        "frame": "first"
                    },
                    {
                        "image": "https://example.com/last.jpg",
                        "frame": "last"
                    }
                ]
            },
            "duration": 10,
            "resolution": "1080p",
            "settings": {
                "multiClip": True,
                "audio": True,
                "thinking": "enabled"
            }
        })


asyncio.run(main())

curl https://api.runware.ai/v1 \
  -H "Authorization: Bearer $RUNWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[
    {
      "taskType": "videoInference",
      "taskUUID": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
      "model": "pixverse:1@8",
      "positivePrompt": "A cinematic multi-shot sequence... Start with... Cut to... Shift to... End with...",
      "inputs": {
        "frameImages": [
          {
            "image": "https://example.com/first.jpg",
            "frame": "first"
          },
          {
            "image": "https://example.com/last.jpg",
            "frame": "last"
          }
        ]
      },
      "duration": 10,
      "resolution": "1080p",
      "settings": {
        "multiClip": true,
        "audio": true,
        "thinking": "enabled"
      }
    }
  ]'

runware run pixverse:1@8 \
  positivePrompt="A cinematic multi-shot sequence... Start with... Cut to... Shift to... End with..." \
  inputs.frameImages.0.image=https://example.com/first.jpg \
  inputs.frameImages.0.frame=first \
  inputs.frameImages.1.image=https://example.com/last.jpg \
  inputs.frameImages.1.frame=last \
  duration=10 \
  resolution=1080p \
  settings.multiClip=true \
  settings.audio=true \
  settings.thinking=enabled

{
  "taskType": "videoInference",
  "taskUUID": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
  "model": "pixverse:1@8",
  "positivePrompt": "A cinematic multi-shot sequence... Start with... Cut to... Shift to... End with...",
  "inputs": {
    "frameImages": [
      {
        "image": "https://example.com/first.jpg",
        "frame": "first"
      },
      {
        "image": "https://example.com/last.jpg",
        "frame": "last"
      }
    ]
  },
  "duration": 10,
  "resolution": "1080p",
  "settings": {
    "multiClip": true,
    "audio": true,
    "thinking": "enabled"
  }
}

[
  {
    "taskType": "videoInference",
    "taskUUID": "a1b2c3d4-e5f6-7890-abcd-ef1234567890",
    "videoUUID": "9c1b2d3a-4e5f-6789-abcd-ef0123456789",
    "videoURL": "https://vm.runware.ai/video/os/a14d18/ws/2/vi/9c1b2d3a-4e5f-6789-abcd-ef0123456789.mp4",
    "cost": 1.15
  }
]

One required field, the rest optional:

• positivePrompt is required and capped at 2048 characters. For multi-shot reels, the prompt drives the shot list. See Writing prompts for cuts below.
• inputs.frameImages accepts up to two anchor frames, each pinned to a position via frame: "first" or frame: "last" (also 0 and -1). Each image accepts a URL, base64 string, data URI, or a UUID from the Image Upload API. The output aspect ratio is derived from the anchors.
• resolution is "360p", "540p", "720p" (default), or "1080p". width and height are not accepted in this mode, the schema forbids them whenever frameImages is present.
• duration is 1 to 15 seconds, default 5.
• settings.multiClip flips the model into multi-shot mode. Off by default. See the next section.
• settings.audio enables synchronized audio. Off by default. See Audio surcharge below.
• settings.thinking controls reasoning depth. See The thinking parameter below.

Three input shapes

When you work with anchor frames, the decision tree is short: how many anchors do you have, and do you want one continuous shot or a cut reel?

The same anchor frame produces a very different output depending on the flag. Below, a single image of a chef plating a dish drives two runs of the same model. The only difference between them is settings.multiClip.

Top-down view of professional chef hands plating a fine dining dish on a wide white ceramic plate, modern restaurant kitchen station with brushed steel surfaces softly defocused behind the plate

The first run reads the anchor as the opening frame of a single take and continues the action inside one continuous shot. The second run reads the anchor as the opening of a sequence and invents the cuts between the kitchen angles. The model holds the plate, the garnish, and the brushed-steel surroundings consistent across each shot, which is the part that traditional re-rolls of separate clips would have broken.

The two-frame interpolation mode lives in Smooth transitions further down, where it has a worked example of its own.

Writing prompts for cuts

multiClip opens the door to cuts, but it does not decide where they happen. The prompt does. V6 honors a beat-structured prompt that reads like a shot list and largely ignores prompts that read like image captions.

The pattern is four phrases anchored to four positions in the runtime:

• Start with for the opening shot
• Cut to for the second shot
• Shift to for the third shot
• End with (or End on) for the closing shot

You can drop one or two of those if the duration is shorter, but the four-beat structure produces the cleanest multi-shot output at 5 to 10 seconds. The model treats each beat as a hard cut and assembles them into a single reel.

The same single anchor image of the boutique handbag below drove two runs with identical settings. Only the prompt structure changed.

A matte black designer leather handbag with gold hardware sitting centred on a polished white marble plinth in a high end boutique window, soft warm overhead spotlight on the bag

The caption prompt described the mood and left the camera choices to the model. The model defaulted to a single slow push with multiClip on, which is the wrong default for a multiClip request. The beat-structured prompt named four explicit shots and got a four-shot reel: the push-in, the side angle, the top-down, the hand-off. The bag itself stays identical across every cut because the anchor frame is doing that work, not the prompt.

Holding the subject across cuts

A reader of the prompt above will notice the closing clause: "consistent bag design across every shot, no dialogue". Naming what must stay continuous is as important as naming the cuts. The model reads the multi-shot prompt and treats unmentioned attributes as fair game. If you do not pin the subject's identity, the bag's clasp can shift between shots, the wine glass can re-render with a different stem, the chef's gloves can change color.

The cue is short and goes at the end of the prompt. "Consistent bag design across every shot." "Preserve the same color grade across cuts." Each clause names one attribute you want preserved.

The vintage Vespa anchor below makes the difference visible. The same single anchor frame drives two runs with the same four-beat prompt. The only difference is whether the prompt closes with a clause naming the scooter's identifying details.

A vintage cherry red Vespa scooter parked on worn cobblestones outside a sun drenched Roman cafe, three quarter angle facing the camera, polished chrome handlebars catching the warm afternoon light, classic round headlight, a small Italian flag sticker on the front fairing

A cinematic multi-shot lifestyle vignette of the vintage cherry red Vespa scooter in Rome. Start with the wide establishing shot from the first frame. Cut to a low ground angle close on the chrome handlebars and the round headlight catching the sun. Shift to an over the shoulder shot looking down the cobblestone alleyway with the scooter in the foreground. End on a rider's hand reaching for the ignition key as they mount the scooter from behind.

A cinematic multi-shot lifestyle vignette of the vintage cherry red Vespa scooter in Rome. Start with the wide establishing shot from the first frame. Cut to a low ground angle close on the chrome handlebars and the round headlight catching the sun. Shift to an over the shoulder shot looking down the cobblestone alleyway with the scooter in the foreground. End on a rider's hand reaching for the ignition key as they mount the scooter from behind. Preserve the exact cherry red bodywork color, the polished chrome handlebar shape, the round single headlight, the worn leather seat, and the small Italian flag sticker on the front fairing across every shot.

The loose run reads the anchor at the open and then lets the small details drift as the camera changes angle. The headlight reshape between the close-up and the over-the-shoulder. The Italian flag sticker is gone by the last shot. The seat texture flips between the cuts. None of this would survive a brand-side review.

The locked run pins the bodywork color, the chrome handlebar shape, the round headlight, the worn leather seat, and the flag sticker in the closing clause. The model carries each one through. The scooter at the end is recognizably the same scooter from the open.

Cuts only render cleanly when the model recognizes the subject as continuous. A prompt that asks for cuts between a sneaker and a city skyline gives the model no continuous thread to carry. Multi-shot V6 works on a single subject moving through a sequence of moments, not a montage of unrelated images.

Smooth transitions

multiClip off with two anchors is a different tool from multiClip on. The model treats the second anchor as the target state and renders one continuous interpolation from the first frame to the last. There are no cuts. Anything that differs between the two anchors morphs smoothly across the runtime.

The cherry blossom pair below demonstrates the shape. The first anchor is a bare branch on a grey overcast sky. The second is the same branch in full pink bloom against the same sky. The prompt names the transition arc and asks for the composition to stay locked.

A single bare cherry blossom branch arcing diagonally across a soft uniform grey overcast sky, dark dormant branch, fine dry bark, no leaves, no buds

A single cherry blossom branch arcing diagonally across a soft uniform grey overcast sky, now in full pink bloom, dense delicate pale pink flowers covering every twig

The output is one take, no cuts. The branch stays in the same position across the eight seconds. Only the flowers evolve, which is the change the anchor pair encoded. Reach for this shape when the change you want is a single visual evolution rather than a sequence of moments. Day to night on the same skyline. Empty stage to packed venue from the same camera angle. Anything where the two endpoints define the whole arc and there is nothing in between worth cutting to.

The same shape works for atmospheric and lighting changes, not just changes in the subject's physical state. A pair of anchors of the same view at two different times of day gives the model the brackets it needs to evolve the light without inventing a new scene. The Manhattan skyline pair below holds the camera position, the building silhouettes, and the bridge composition locked across both anchors. Only the time of day changes.

A wide cinematic skyline view of Manhattan from across the East River at midday, crisp blue sky with scattered white clouds, sharp midday sunlight on the glass facades of the towers, the Brooklyn Bridge cables crossing the lower left foreground

A wide cinematic skyline view of Manhattan from across the East River at deep night, dark navy sky, warm pinpoint glow of thousands of lit office windows scattered across the towers, the Brooklyn Bridge illuminated with its cable lights

The buildings stay where they were. The bridge cables stay where they were. The lighting moves through the full arc from noon glare to dusk to a dense field of warm window lights. This is the same mode the cherry blossom used. The difference is only what the anchors encode: the cherry blossom anchors locked composition and changed the subject's state, the skyline anchors lock composition and change the lighting.

The thinking parameter

settings.thinking controls how much reasoning the model spends interpreting the prompt and the anchors. The values it accepts depend on how many anchor frames the request carries:

• With zero or one anchor frame, thinking takes "enabled", "disabled", or "auto" (default).
• With two anchor frames, thinking narrows to "enabled" or "disabled". The "auto" value is rejected, you have to make the call explicitly.

Across both shapes the meaning is the same:

• "enabled" spends extra compute on prompt interpretation, shot planning, and consistency tracking. Slower per call, more faithful to long prompts.
• "disabled" skips the reasoning step. Faster, weaker adherence to complex shot lists.
• "auto" lets the model decide.

For single-anchor reels, leaving it on "auto" is the right default. For two-anchor calls you have to pick, and "enabled" is the safer choice when the prompt names beats or the anchor pair encodes a complex arc. Drop to "disabled" for fast iteration when you are still tuning a prompt and the output quality is secondary.

Audio surcharge

settings.audio: true generates a synchronized audio track at the same time as the video. The track is environmental by default: footsteps, ambient room tone, weather, mechanical sounds, distant voices. It is not a substitute for narration or licensed music.

The surcharge is real. From the model's pricing tiers:

The premium earns its place on atmospheric reels where the soundtrack is the environment itself: wind through vines, footsteps on cobblestones, water in a fountain, the hum of a kitchen pass. PixVerse renders these in sync with the on-screen motion, which an editor would otherwise time and mix by hand. The synced bed also carries across cuts, so the audio does not pop between shots in a multiClip reel.

Leave it off when post will replace the track anyway. A reel headed into a longer edit with its own music or voiceover gains nothing from a synced ambient bed underneath. Same for assets destined for silent-autoplay platforms where the audio never reaches the audience. Dialog and licensed music belong in the post pipeline regardless of this flag.

Tips

1. Use the beat pattern for any multiClip request. Start with / Cut to / Shift to / End with is the structure the model honors. Captioned mood prompts produce single takes even with multiClip: true, which is the wrong output for the parameter.

2. Always close the prompt with a consistency clause. Name the subject attributes the model must hold across cuts. Without that clause, hardware drifts, color grades reset between shots, and characters lose identifying details from one beat to the next.

3. Match the anchor framing to its pinned position. A frame: "first" anchor should look like the opening of the reel, a frame: "last" anchor like the closing. Mismatched composition between the anchors and their pinned moments forces the model to redraw the bracketing frames, which defeats the point of anchoring.

4. Two anchors with multiClip: false is a transition, not a reel. Reach for it when the change is a single visual arc with no cuts worth making. The cherry blossom shape, not the vineyard shape.

5. thinking: "auto" for single-anchor reels, pick explicitly for two-anchor. The validator accepts "auto" only with zero or one frame. Two-anchor calls require either "enabled" or "disabled". Reach for "enabled" when the result missed a beat, dropped a consistency cue, or skipped a transition step.

6. Pay the audio surcharge for atmospheric reels, skip it for cuts that will get a soundtrack downstream. Synced environmental audio is hard to add in post and easy to lose. Music and dialog belong in the edit, not in the generation call.