VibePrevisBench: Evaluating Agentic Coding for 4D Scene Generation in Blender
Across 11 coding agents, inspecting their own Blender renders did not reliably lead to correcting errors.
VibePrevisBench tests agents on 242 tasks that turn scripts, photos, continuations and reference videos into executable 4D scenes and rendered proxy videos. It checks both the scene’s underlying facts and whether those facts are visible in the render. The authors report that basic camera work and staging appear largely mastered, while harder requests expose weaknesses in spatial understanding and camera motion. They call for explicit checks on difficult requirements and item-by-item verification against the brief.
HF Daily Papers' note
VibePrevisBench tests agents on 242 tasks that turn scripts, photos, continuations and reference videos into executable 4D scenes and rendered proxy videos. It checks both the scene’s underlying facts and whether those facts are visible in the render. The authors report that basic camera work and staging appear largely mastered, while harder requests expose weaknesses in spatial understanding and camera motion. They call for explicit checks on difficult requirements and item-by-item verification against the brief.
HF Daily Papers' note
score 4