Previz
Block a shot with simple shapes and a keyed camera in Blender, render it as a clay video, and let Seedance 2.5 or another model follow it.
Previz means blocking the shot before you film it: where the camera goes, where people stand, when they move. For AI video you block it in Blender with plain shapes, render a short "clay video" from the camera, and give that video to the model as a reference. It is the most repeated 3D pattern among creators, and ByteDance now documents it officially for Seedance 2.5.
Expect close, not exact. The first creator to show this pattern called the tracking close but not perfect.
What ByteDance says
The Seedance 2.5 prompt guide calls this "3D clay model video reference". The official points:
- Two levels. Coarse: a clay video carries the motion, camera moves, paths and lighting changes, while image references carry the look of subjects, scenes and props. Fine: a fully modelled scene that the model re-renders.
- Simple shapes work better. ByteDance recommends plain geometric primitives for people, objects and animals in the current version.
- No overlays. Remove motion path lines, axes, camera cones and other overlays. In practice: render from the camera, do not screen record the viewport.
- Say what to use. Tell the model which parts of the clay video to take: camera only, camera and motion, or also lighting.
- Colour codes. Give each stand in its own colour and map each colour to an image reference in the prompt. Still describe the shot in text, consistent with the clay video.
- Clay video plus keyframes. ByteDance's own 30 second example uses the clay video only for camera, rhythm, blocking and character animation, and takes the look from 10 keyframe images.
- Storyboards are loose. A storyboard grid is followed loosely. For strict alignment use keyframe images in order.
ByteDance also publishes a prompt optimiser skill for coding agents:
npx --yes skills@latest add "https://arkdocs-en.tos-ap-southeast-1.volces.com/skills/" --skill sd25-pe --yesThen run /sd25-pe followed by your prompt.
Reference video limits (Seedance 2.5)
| Limit | Value |
|---|---|
| Number of videos | Up to 10, 30 s in total, each 2 to 30 s |
| Frame rate | 24 to 60 fps |
| Pixels per frame | At least 407,696. 854x480 passes, 640x360 fails |
| Sides and shape | 300 to 6000 px per side, aspect ratio 0.4 to 2.5 |
| File | mp4 or mov, up to 200 MB |
| People | No real human faces in references |
What creators learned
Keep the blockout barebones
This is the consensus. In a same prompt test, a barebones blockout beat a detailed one, because Seedance copies detail literally. Posed human models gave stiff motion, and plain blocks with image references worked better. Rigid 3D animation carried over even after prompting around it. One creator removed weapons from the blockout and kept them only in the character sheet. Some of these posts are sponsored (CapCut and PixVerse creator programs), but organic tests agree, and so does the official note.
Use linear keys and small camera moves
Seedance copies camera dynamics, not just angles. One detailed creator found big orbits and eased moves competed with the action, so the final cameras used about 20 degrees of movement, linear keys and constant speed. The same creator noticed that moves which look right in Blender usually come out too fast, so slow them down. This is a single source, but the most detailed one.
Give each reference one job
The same creator used one Seedance 2.5 generation with the blockout for the camera and a real dance clip for the body. The stand in (a cone) never moved, so it did not fight the dance. Location plates were rendered to match the blockout angles, with 4 character views and 3 location views. For dialogue, stand ins got eyes, a nose and a moving mouth so the model knew who faced where and who was talking.
Moderation still applies
Seedance refused a clay preview on copyright grounds for one user. Runway moderates violence hard, so one fight was restaged without the sword. Real faces are not accepted as Seedance 2.5 references.
Step by step
Write the brief
Give Opus the shot length, frame rate, aspect ratio and output resolution, and ask for real world scale in metres. See the greybox prompt on Prompts.
Greybox the set
Ground, walls and large set pieces as primitives. Characters as colour coded capsules or mannequins with a small nose block to show facing. Add a simple face only if the shot has dialogue. No detail, no textures.
Key the camera
One camera per shot. Set focal length and height, then 2 to 4 keys. Use linear keys, constant speed and modest moves when the subject carries the action. Save big orbits for empty establishing shots.
Animate marks, not limbs
Move stand ins only for start, stop, turn and exit. If the body motion comes from another reference (a dance clip, motion capture, Kling Motion Control), keep the stand in still.
Check the camera view
Ask for camera view screenshots at the first, middle and last frame, and a list of frames where a stand in leaves the frame or cuts through geometry.
Render the clay video
Render from the camera with flat or Workbench colours, no overlays, at the exact clip length. Settings are on Hand off.
Add stills (optional)
Render matching location plates and the first and last frames from the same cameras, then restyle them with an image model.
Write the video prompt from the scene
Have Opus read the scene and write the prompt: one job per input, each colour mapped to a reference, the final look described, and a clear line not to copy the clay look.
Draft cheap, finish once
Draft with Seedance Draft mode or at 480p, pick the take, then render the final once.
Render, do not screen record
A screen recording of the viewport carries grids, gizmos and camera outlines into the reference, and ByteDance asks you to remove them. Render from the scene camera every time.
Example prompt for Seedance 2.5
This follows the official reference syntax. Upload the clay video first, then the images in the order listed.
@Video 1 is a clay previz render. Use only its camera path, framing, the positions of
the shapes and the timing. Do not copy its flat grey shading, capsule shapes or blank ground.
@Image 1 is the runner: the orange capsule in @Video 1 is this character.
@Image 2 is the chaser: the blue capsule in @Video 1 is this character.
@Image 3 is the rooftop location and its night lighting.
0 to 4 seconds: the runner sprints across a wet rooftop toward the parapet, camera tracking
behind him at shoulder height.
4 to 6 seconds: he jumps the gap between buildings, city lights below.
6 to 8 seconds: the chaser arrives and stops at the parapet, breathing hard.
<rain on concrete> <distant sirens> <footsteps splashing>
No subtitles. No BGM.Shortcuts and tools
Several vendors sell this pattern as a product: a Dreamina "Clay Renderer" plugin, a Higgsfield add-on that runs Seedance 2.5 inside Blender, a Scenario plugin, and ComfyUI workflows. Most of the posts about them are paid campaigns; see Troubleshooting. Open source options include Motion Previs Studio, which pulls pose, depth and camera from reference footage and has its own MCP server. If you want a handheld camera without Blender skills, one creator used an iPhone as a virtual camera inside the Blender set (VirtuCamera 2, stated as not sponsored).
Sources
- Seedance 2.5 prompt guide
- Seedance 2.5 tutorial (reference limits)
- Barebones test: GrandVisionsAI (organic)
- Camera dynamics and reference jobs: gensoFX thread, camera notes, dialogue stand ins (organic)
- Literal copying: smoermeow (organic), kojomojoworld (sponsored)
- Origin of the pattern: reidhannaford (organic)
- Full worked prompt: Ror_Fly (organic)
- Motion Previs Studio