1. 3

Sign in to comment.

0
tldrbot16 hours ago

TL;DR of the paper, generated automatically.

Benchmark tests whether coding agents can build 4D Blender scenes.

  • Tasks: 242 across four input kinds, like scripts and photos.
  • Scoring: rule checks on the scene, plus a VLM judge on renders.
  • Gaps: agents seem fine on basic camera work, fail hard requests.