Video models are good at look and bad at layout. Ask one to invent the camera move, the staging and the scale of a room all at once and you get a lot of rejected takes. So before a shot goes to the video model, I block it out in Blender, and the model only has to supply the look.
Here’s the process in a short video:
The short version:
- Scratch dialogue first. Every line of the scene is generated as rough audio, so the timing of the whole scene comes from the voices rather than guesswork.
- A rough set. Plain boxes and primitives, built to true scale in meters. Nothing here is meant to look finished.
- Capsule characters and primitive props. Stand-ins that carry position, height and motion, and nothing else.
- Rough animation timed to the audio. Characters walk, turn and react on the beats of the dialogue.
- One camera per shot, each bound to a timeline marker, so a single
.blendfile holds the whole scene and I can adjust it shot by shot. - Render each shot to a short video that carries the camera, layout, scale and motion.
That video goes to Seedance along with a set plate and the character references. The blockout supplies the where and the when, and the references supply the what it looks like. The blockout has to stay crude on purpose: the model copies anything in the video that looks finished, so plain shapes keep it from copying the stand-ins.
It’s more work up front, but it cuts down on retakes, and a lot of the cold open was made this way.