VibePrevisBench tests coding agents on 4D Blender scenes
The point
VibePrevisBench tests coding agents on 4D scene generation in Blender. An agent turns a brief into an executable 4D scene and renders a proxy video. The benchmark covers 242 tasks across 11 frontier coding agents.
Why it matters
For builders, code that runs still fails if the render misses space and camera intent.
Key details
- 242 tasks span four inputs: scripts, photos, continuations, and reference videos
- Scene Check compares exported scenes with hidden ground truth in nine dimensions
- Agents handle basic camera work but fail to correct errors in their renders
Original ~6 min → 25 sec here Source: arXiv cs.AI, Nov 13, 2026