VibePrevisBench: 242-task benchmark tests coding agents on 4D Blender scene generation
Kling-Keye · hf · 2026-11-16
VibePrevisBench from Tianjin University, Kling AI and partners is the first benchmark evaluating coding agents that reconstruct 4D scenes in Blender to produce graybox proxy videos for controllable video generation.
- Scale: 242 tasks across four input types (scripts, photos, continuations, reference videos)
- Dual scoring: rule-based Scene Check (re-runs code, compares against hidden ground truth on 9 dimensions) plus a VLM-based Screen Check
- Subjects: 11 frontier coding agents
Key findings: multimodal understanding separates agents, especially spatial and camera-motion perception; all agents master basic camera work but fail on challenging requests; agents inspect their own renders yet fail to correct errors. The authors call for better spatial visual understanding, explicit checks, and per-item verification against the brief.
More from coding & agent
- raylib author threatens to block users over AI-generated GitHub issues — yacineMTB · 2026-11-16
- Dev open-sources local Apple Health MCP server that feeds workouts to AI agents and pushes plans back to Apple Watch — PrinzPwnage · 2026-11-16
- Coding agents crack a 1994 Jeep's ECU via a $3 Arduino and start controlling it — LeviTurk · 2026-11-16
- 'My favorite PRs': raylib author hails a change that adds features by deleting 100+ lines — yacineMTB · 2026-11-16
- Indie dev shows agent workflow turning before/after UI screenshots into beat-synced changelog videos — yihui_indie · 2026-11-16
- jscdec: Pure-JS V8 bytecode decompiler ships with Electron .asar/.exe unpacking — cyb3rops · 2026-11-16