News

VibePrevisBench: Benchmarking 4D Scene Generation in Blender

VibePrevisBench introduces a benchmark for evaluating coding agents' ability to generate 4D scenes in Blender, featuring 242 tasks and two evaluation metrics.

VibePrevisBench: Benchmarking 4D Scene Generation in Blender

On November 17, 2026, researchers led by Yuzhe Chen, Muxi Diao, and others introduced VibePrevisBench, a novel benchmark for evaluating coding agents' ability to generate 4D scenes in Blender. Published on arXiv (2611.15789) and highlighted on HuggingFace Daily Papers, the project aims to assess how effectively agents can translate briefs into executable 4D scenes and corresponding proxy videos.

Key Features of VibePrevisBench

VibePrevisBench evaluates coding agents across 242 tasks, categorized into four input types: scripts (94 tasks), photos (39 tasks), continuations (23 tasks), and reference videos (86 tasks). These tasks span seven scene domains and 33 locations, providing a comprehensive testbed for agentic coding capabilities.

The benchmark employs two evaluation metrics:

  • Rule-based Scene Check: This metric re-runs the agent-generated code and compares the exported 4D scene against hidden ground truth (GT) facts across nine dimensions, including spatial layout, actions, cameras, and physics.
  • VLM-based Screen Check: A vision-language model (VLM) judge assesses whether the correct GT facts are visually recognizable in the proxy video.

In total, the benchmark incorporates 84,406 scorable hidden GT facts, ensuring rigorous evaluation of agent performance.

Findings from Agent Evaluations

The research team analyzed the performance of 11 frontier coding agents, uncovering three key insights:

  1. Multimodal understanding, particularly in spatial and camera motion comprehension, differentiates agent performance.
  2. While agents have mastered basic camera work and staging, they struggle with complex and challenging requests.
  3. Agents often inspect their own renders but fail to correct identified errors, highlighting a need for more robust item-by-item verification processes.

Open-Source Implementation and Accessibility

The VibePrevisBench project is fully open-source, with its code and data released on November 11, 2026. The GitHub repository (vibeprevisbench/vibeprevisbench) includes the complete evaluation pipeline, allowing users to test agents with a one-command evaluation tool, vibeprevisbench-eval. Additionally, the project homepage features benchmark results, an interactive leaderboard, and 14 case studies.

By providing a robust framework for evaluating agentic coding in 4D scene generation, VibePrevisBench sets a new standard for advancing AI-driven video generation technologies.

Sign in to save this article Free account: build your research collection and return to read it anytime.