Stop Tweaking Prompts—Someone Used Blender Blockout Figures to Lock Down the Camera in AI Video

Seedance 2.5 generated a 4K video of a man in armor running, switching worlds every few seconds: a medieval village, a bamboo forest, Mars, a neon city in the rain, a snow village, ancient Egypt, 1920s streets, a jungle, a clay village, ancient Rome, and a sunset beach.

But what’s really worth watching is the layer of drab gray blockout figures below the frame.

Blender blockout—often called rough blocking or layout blocking in Chinese—has been used in animation and VFX for decades: use spheres, cubes, and simple human figures to establish character positions and scene proportions, determine how the camera moves, and leave materials, lighting, and detail for later. It’s cheap, fast, and changing a shot only takes a few minutes.

In AI video generation, the most disobedient variable is the camera. If you write “the camera follows the running man,” the model may give you a random combination of push, pull, pan, and tilt; the character deforms in motion, and the background drifts. No matter how detailed the prompt is, the camera trajectory is still a lottery.

This demo’s approach is to render the blockout as a reference video and feed it to Seedance 2.5. The model doesn’t need to guess how the camera moves; it only needs to generate the visual content along a set trajectory. The camera is locked down, and the rest is left to the string of scene changes in the prompt.

The prompt itself is also worth a look: a man runs, the world changes every few seconds, 11 scenes are listed, and the last line is “follow reference camera and motion.” The key is in that last line.

In the replies, Proofsmith said: “4k seedance with a blender blockout as the base is a clever way to lock the camera.” That observation is far more useful than the “Fantastic Share” comments below.

The value of this demo isn’t the 4K, nor how flashy those 11 scenes are. It offers an engineering path: use the most basic blockout from the traditional 3D pipeline to constrain the most disobedient variable in generative models. Tweaking prompts is negotiating with the model; blockout draws the tracks for it directly.

For people who already know Blender, this combination has almost no learning cost. For people who don’t, this may be the most practical reason to learn blockout.