Ray3.2

Introduction

Modify Video re-imagines footage you already have. You pass a clip and a prompt, and Ray 3.2 re-renders it guided by the source, carrying the original motion and timing through into a new look. It's a transformation tool, not a surgical editor: the prompt reinterprets the whole frame, so faces and backgrounds are reimagined along with the style rather than held pixel-for-pixel.

The playing carries straight through while everything else is rebuilt from the prompt, and that is the model's sweet spot: bold transformations where the performance is the throughline and the new look is the point. Editing keeps the source aspect ratio, so changing format is a separate reframing step. This guide covers the request, the strength dial that sets how far the result departs from the source, and the per-signal controls.

The request

An edit is a normal video request with a source clip in inputs.video and a prompt describing the look you want. The model follows the source's motion and composition, and the prompt drives everything else.

[
  {
    "taskType": "videoInference",
    "taskUUID": "e5d6c7b8-9a0b-1c2d-3e4f-5a6b7c8d9e0f",
    "model": "luma:ray@3.2",
    "positivePrompt": "Transform the violinist into a luminous figure of flowing golden light against a dark void, ribbons of light streaming from the bow and tracing his body, following his exact bowing and motion.",
    "inputs": {
      "video": "https://example.com/violinist.mp4"
    },
    "settings": {
      "edit": { "strength": "flex_2" }
    }
  }
]

{
  "data": [
    {
      "taskType": "videoInference",
      "taskUUID": "e5d6c7b8-9a0b-1c2d-3e4f-5a6b7c8d9e0f",
      "videoUUID": "f6e5d4c3-b2a1-0987-6543-21fedcba0987",
      "videoURL": "https://vm.runware.ai/video/os/a14d18/ws/2/vi/f6e5d4c3-b2a1-0987-6543-21fedcba0987.mp4"
    }
  ]
}

inputs.video takes a URL or the UUID of a previous generation. Describe the target look and let the motion carry. The model rebuilds the rest of the frame from there.

Controlling how much changes

settings.edit.strength is the preserve-versus-reinterpret dial. It runs in three bands: adhere stays close to the source, flex takes moderate liberties, and reimagine treats the source as loose guidance. Each band has three levels (adhere_1 through reimagine_3) for finer steps.

Below, one prompt turns the dolphin into a cyberpunk cybernetic creature, run at one level from each band against the source it started from:

Each band sets how far the result departs from the source. At adhere_1 the cyber-dolphin stays locked to the real leap and the natural ocean. At reimagine_3 the model reworks the creature and scene more freely, keeping only a loose echo of the source.

"settings": {
  "edit": { "strength": "reimagine_3" }
}

Per-signal controls

For tighter command than a single dial, settings.edit.controls exposes the individual conditioning signals:

• poseStrength (precise or coarse) sets how strictly the subject's skeleton is followed.
• depthBlur and normalsAugmentation (both 0 to 1) loosen scene geometry and surface detail as they rise, granting more reinterpretation.
• trajectorySparsity (0 to 1) controls how many motion anchors hold the movement in place.
• face biases the result toward a recognizable face, though a heavy transform still reinterprets it.

The transformation below pins the dolphin's exact leap with poseStrength while applying the same cyberpunk transform:

"settings": {
  "edit": {
    "controls": { "poseStrength": "precise" }
  }
}

Changing the setting

The same mechanism transforms the world around a shot. The source street below is reskinned into deep winter, the camera move carried through. As with any edit the buildings are rebuilt rather than copied, so reach for this when you want a fresh take on a setting, not the old one left untouched under new weather.

Relighting a scene or restyling the subject runs the same way: describe the new look, then pick a strength for how far the result should travel from the source.

Tips

1. Describe the look you want, not a precise edit. Name the target style and let the motion carry. The model re-renders the rest, so don't count on faces or backgrounds staying untouched.

2. Start low on the strength dial. adhere and flex keep the shape and motion close. Reach for reimagine only when you want the model to rework the subject itself.

3. Turn on face for people. It biases the result toward a recognizable face, though a heavy transform will still reinterpret it.

4. Let autoControls handle the first pass. It derives the conditioning from the source, which is usually enough before you reach for manual signals.

5. Reframe in a separate step. Editing holds the source aspect ratio, so changing format is its own reframing request rather than part of the edit.