A permanent step for concepting any narrative video with a recurring location. Build the set once in 3D, block every camera against it, and let AI only dress the shots — so the room never drifts between frames.


projects/big-gorilla/set_spec.json.cameras entry per storyboard beat (loc, target, lens) — the reusable, exact version of marking angles on a floorplan. Subject stand-ins can appear only in the beats that need them.Blender -b -P tools/previs/greybox_kit.py -- <spec.json> <camera> <out_dir> → a color greybox + a depth map.| Engine | Strength | Use for |
|---|---|---|
| FLUX Kontext | Photoreal + holds layout + real subjects | Default |
| depth-ControlNet | Pixel-locked geometry | Hard continuity (+ shaped stand-in) |
| Gemini nano-banana | Holds layout, matte look | Backup / variety |
| gpt-image-1 | Best polish, drifts | Look-dev only |
Invoke the /set-designer skill at the start of a concept, or follow .claude/skills/set-designer/SKILL.md. Toolkit lives in tools/previs/; dressing runners in tools/ai_video/. Proven on the Big Gorilla spot (see its Continuity Proof and Model Bake-off pages).