Video (from scene)
Use the keyframes you set in Compose to drive an AI video model's motion, or render the scene as video without AI via Direct Render.
You authored keyframes on the shot in Compose. Video (from scene) takes those keyframes as the motion source for the video. The model follows the camera move, the subject motion, and the pose changes you set, rather than inventing motion of its own. Or pick Direct Render and skip the AI entirely.
What it does
Most video models start from a still image and a prompt and invent the motion. Video (from scene) is different: it takes the animated 3D scene as the input. The keyframes you set in Compose drive what the model renders. The output follows the camera path, the character pose changes, and the object motion you authored, within whatever the chosen model can interpret.
This is the mode to reach for when the motion matters more than the look. If you've spent time animating a precise camera move or character action, this is how you preserve it through video generation.
How to use it
Author the animation in Compose. Set keyframes on the camera, subjects, and any animated objects in the shot. See Animate an object for the full flow.
Switch to Visualize. The shot's animation is preserved across modes.
Set the mode dropdown to Video (from scene). The panel updates to show the model picker for this mode.
Pick a model. Three options, each does something different (see below).
Write a prompt (skip this for Direct Render; the scene drives everything).
Click Capture Animation. The job runs server-side. The output lands as the shot's video.
The three model options
Seedance 2.0 (ByteDance). Generative video that follows the scene's motion. Use it when you want a stylized AI output that respects your authored blocking.
Luma Ray 3.14 (Luma). Video-to-video model with first-and-last frame anchoring. Tightest motion fidelity of the AI options.
Direct Render (Intangible). Non-AI. Renders the 3D scene as video, frame-for-frame, no diffusion model involved. See Direct Render for the full picture.
Motion versus appearance
When you use Video (from scene) with an AI model, the motion is driven by your keyframes but the appearance still comes from the prompt and any reference images. A dolly through an empty 3D scene produces a dolly through whatever the prompt describes. A subject animating across frame stays animated, but its visual identity is decided by the prompt.
If you want the appearance locked to your scene too, that's Direct Render.
When to reach for each option
Stylized AI video that follows your authored motion
Seedance 2.0
Tighter motion fidelity with start / end anchoring
Luma Ray 3.14
Wireframe preview of the authored shot, no AI, no credit
A still-image generation with no motion
Switch the mode dropdown back to Image
Video from a single start image, no scene animation
Switch to Video (from keyframes), see Generate video
Video from two anchor images
Switch to Video (from keyframes), see First and last frame
Limits and known issues
The model interprets your motion; it does not replay it. Seedance 2.0 and Luma Ray 3.14 honor authored motion, but they still resample through their own video generation. Sharp keyframe changes can soften; fast camera moves can drift. If a move has to land exactly, use Direct Render.
Direct Render output is wireframe / greybox, not photoreal. It captures the 3D scene as it appears in the viewport. For a finished look, render through one of the AI models.
No audio. The audio toggle that appears on other video modes is not present here. Add audio later in your editing tool, or use Video (from keyframes) with a model that generates audio.
Related
Direct Render – the non-AI option inside this mode.
Animate an object – how to author the motion that drives this.
Animation and shot time – the concept.
Generate video – the mode menu and the keyframe-driven video path.
Models – the full model reference.
Last updated
Was this helpful?