Greybox to AI video
Grey boxes, one character, a camera. That is the whole input. The model fills in the concrete, the coat and the weather, and keeps the shot you blocked.
Greybox left of the wipe, Seedance 2.5 on the right. Same camera on every cut.
Top: the blockout. Bottom: the generation. The pillar, the fall and the framing all survive.
What a greybox is
A greybox (greybox, blockout, white model, clay render — same thing) is a 3D scene built from untextured primitives: cubes for buildings, a cylinder for a lamp post, a flat plane for the street, a stand-in figure for the actor. Game studios have blocked levels this way for decades, and film previs departments do the same thing with slightly nicer boxes. Nothing in it is meant to look final. Everything in it is meant to be measured: where the walls are, how tall the character is, where the camera sits and what lens it has.
That last property is why it matters for AI video. A prompt says low angle; a greybox says the camera is 0.6 m off the ground, 2.4 m from the subject, on a 24 mm lens, and it says it in pixels the model can see.
Why video models take a greybox better than a camera prompt
Current video models accept two kinds of visual input alongside text: a first frame (image-to-video) and, on some models, a reference clip (video-to-video or reference-guided generation). Both are pixel channels. The model does not parse them into "camera at 0.6 m"; it matches them. A clean grey render gives it composition, blocking and camera motion with nothing else to latch onto, so what comes back keeps the shot and invents only the surfaces.
A textured render or a photo does the opposite. The model tries to keep the wood grain and the shirt too, and when your reference was a rough placeholder that is exactly what you did not want preserved. Grey is not a limitation of the workflow. Grey is the point.
Keep the export flat: no outlines, no gizmos, no grid, no text overlays. Anything drawn on the frame is something the model may try to keep. A blue selection box on the ground plane can come back as a blue crate.
The block-then-generate loop
- Block the set. Boxes for the architecture, a plane for the ground, a stand-in for each character. Scale matters more than detail: doors at 2 m, a car at 4.5 m long.
- Place the camera. Angle, height, lens, then the move: a dolly rail, a crane, a cut at frame 40. This is the part a prompt cannot hold.
- Export a frame or a clip. A still for image-to-video models; a short clip for models that take a reference video. Flat shading, the delivery aspect ratio, nothing drawn on top.
- Generate with a short prompt. The prompt now only has to describe materials, light and mood. Camera and blocking are already in the reference.
- Iterate on the greybox, not the prompt. If the shot is wrong, move the camera two metres and export again. That costs nothing. Re-rolling a prompt costs a full generation.
The economics are the reason to bother. Seedance 2.5's list price on BytePlus works out to roughly $17 for 30 seconds of 1080p, and a take with the wrong angle is a take you pay for again. Blocking the shot first turns a paid guess into a free one.
Which models, which channel
| Model | Takes a first frame | Takes a reference clip | What to send |
|---|---|---|---|
| Seedance 2.5 | Yes | Yes | A flat greybox clip for camera moves; a still for locked-off shots. See Seedance camera control. |
| Kling | Yes | Motion-control inputs for character motion; camera via first frame and prompt | A greybox still per shot; keep moves simple. |
| Veo | Yes | No | A greybox still per shot. Describe the move in the prompt; expect drift on long moves. |
Model capabilities change monthly; this table reflects what each accepts as of September 2026. The workflow does not change: give the model pixels, not adjectives.
Do it in the browser
CozyClay is open source previs that runs in a Chromium tab. It exists for exactly this loop: block,
camera, export, generate. The playground on the home page walks through the camera
part in seven moves on the same city block you see in the reels, and npx cozyclay --scene city-block
opens that set on your machine.
FAQ
Greybox, graybox, blockout, white model: are these different?
No. Games say greybox or blockout, film previs says white model or clay, ByteDance's own Seedance material says clay reference. All mean an untextured, flat-shaded 3D scene used to decide layout and camera before final assets exist.
Do I need Blender for this?
No. Any tool that can place boxes, a figure and a camera and export a flat render works. Blender does it; so does CozyClay in a browser with nothing installed. The model only sees the export.
Why not just prompt "low angle, slow dolly in"?
It works as a nudge and fails as a spec. Words carry no height, distance or focal length, so every run fills them in differently. A greybox frame fixes them. See camera control for AI video for the long version.
Should the greybox character be posed?
Yes, roughly. Where the figure stands and which way it faces set eyelines and blocking. Fine pose detail is unnecessary; the model handles limbs. For character motion, a short greybox clip with the figure moving is enough for models that accept reference video.