Seedance 2.5 camera control
Seedance will give you a low angle. It will not give you the same low angle twice unless you show it one. Here is the reference-clip workflow, what to put in the clip, and what to leave out.
Top: a greybox blocked in CozyClay, exported flat. Bottom: Seedance 2.5 from that clip. The pillar, the fall and the camera all hold.
What Seedance 2.5 accepts
Alongside the text prompt, Seedance 2.5 takes a first frame (image-to-video) and reference material: images and short clips the generation should follow. ByteDance's own examples include clay and white-model references, which is the previs use case exactly. The reference is a pixel channel. The model matches what it sees; it does not read a camera description out of it.
That is why a camera prompt drifts and a reference clip does not. "Slow dolly in, low angle" has no numbers in it. A 3-second greybox clip of a camera dollying in at 0.6 m has all of them, in the only form the model understands.
What a missed take costs
On BytePlus ModelArk, Seedance 2.5 at 1080p lists at about $2.84 per 5 seconds, so a 30-second sequence is roughly $17. A take that comes back with the wrong angle is a take you buy again, and camera prompts miss often enough that a six-shot sequence can cost two or three times its list price before it cuts together. Fixing the camera in a reference clip moves the misses to the free part of the process.
The workflow
- Block the set in grey. Boxes and a stand-in figure, at real scale. No textures.
- Set the shot. Camera height, distance, lens, then the move: a dolly rail, a crane, a hold. Decide where the cut lands.
- Export a flat clip. Same aspect ratio as the delivery (16:9 or 9:16), flat shading, 3 to 5 seconds, nothing drawn on top.
- Attach it as the reference. Prompt only for what the clip does not carry: materials, weather, wardrobe, mood. Do not describe the camera again in words; the clip already did.
- Judge the camera, not the render. If the framing is off, move the camera in the blockout and export again. That is free. Then spend on the generation.
Strip the export. Outlines, gizmos, the ground grid, a selection highlight, a frame counter: any of these can come back in the generation as an object. A flat clay pass with a plain sky is the cleanest reference Seedance can get.
Reference clip versus first frame
| You need | Send | Why |
|---|---|---|
| A locked-off shot | First frame (still) | Composition and lens are fixed by the image; nothing moves, so nothing drifts. |
| A camera move | Reference clip | A still cannot carry a dolly or a crane. The clip carries the whole move, frame by frame. |
| Character motion plus camera | Reference clip with the figure moving | The model follows both. Keep the figure grey so it invents the person instead of preserving a mannequin. |
| A cut between two setups | Two references, two generations | Generate per shot and cut in an editor; models do not hold a cut inside one generation reliably. |
Making the clip
Any 3D tool that renders flat can make the reference. Blender works with a clay material and a
workbench render. CozyClay, open source and in the browser, exports the flat clip directly with the
delivery ratio and a camera timeline built in; the reel at the top of this page came out of it.
The home-page playground walks through blocking a shot and drawing a dolly rail in
seven moves, and npx cozyclay --scene city-block opens that set locally.
FAQ
Does Seedance 2.5 support camera control prompts?
It responds to camera language, but as a tendency rather than a specification. The same prompt produces a different angle and distance on each run. A reference clip or a first frame is how you make the camera repeatable.
How long should the reference clip be?
Match the shot you want, typically 3 to 5 seconds. Longer references cost more to generate against and drift more toward the end of the clip.
Does this work on Kling or Veo?
Kling and Veo take a first frame, so locked-off shots transfer directly. Reference-clip camera moves are a Seedance strength; on the others, keep moves short and describe them in the prompt. See camera control for AI video.
Should the greybox have textures or colours?
No. Grey or flat clay. Colour is a cue the model preserves, and you want it inventing the surfaces, not matching placeholders. A single highlight colour on the subject is the most you should add.