# VibePrevisBench: Kling Team Benchmarks Coding Agents on Blender 4D Scenes

> Source: <https://explainx.ai/blog/vibeprevisbench-kling-blender-agent-4d-scene-benchmark-2026?utm_source=rss&utm_medium=feed&utm_campaign=blog>
> Published: 2026-11-16 00:00:00+00:00

**VibePrevisBench tests whether coding agents can turn director briefs into **4D Blender scenes** and **graybox proxy videos** that controllable video models paint over—242 tasks, dual scoring, and public case replays at [vibeprevisbench.github.io](https://vibeprevisbench.github.io/?ref=explainx).** The [arXiv paper](https://arxiv.org/abs/2611.15789?ref=explainx) (2611.15789) comes from researchers affiliated with **Kling AI**, Tianjin University, BUPT, and other institutions. It fills a gap between **SWE-bench-style code fixes** and **Seedance-class video generation**: the intermediate **previsualization** layer.

If your team ships **AI video pipelines**, the benchmark answers a practical question: can your agent **block a multi-shot scene** well enough that [Seedance 2.5](https://explainx.ai/blog/seedance-2-5-bytedance-30-second-4k-ai-video-2026) or similar models inherit correct camera and timing? The authors argue errors in the proxy propagate to final pixels—a point illustrated on the site with **GPT-6 Astra** grayboxes fed to **Seedance 2.5**.

## TL;DR

| Question | Answer | 
|---|---|
| Tasks | **242** across script, photo, continuation, reference video | 
| Delivery | Blender **.blend** + rendered**graybox proxy video** | 
| Gate 1 | **Scene Check** — deterministic facts in 9 dimensions | 
| Gate 2 | **Screen Check** — VLM judge on visibility and story | 
| Agents tested | **11** frontier coding agents (site leaderboard) | 
| Demo stack | Grayboxes by **GPT-6 Astra** ; paint pass by**Seedance 2.5** | 
| Primary site | [vibeprevisbench.github.io](https://vibeprevisbench.github.io/?ref=explainx) | 

## Where this sits in the video stack

Modern controllable video workflows separate **layout** from **appearance**:

1. A **coding or layout agent** builds geometry, character blocking, and camera cuts in Blender.
2. The agent renders a **graybox proxy** (depth-like or flat-shaded preview).
3. A **video diffusion model** conditions on that proxy to output photoreal frames.

explainx.ai has tracked the consumer side in [Seedance 2.5 pricing claims](https://explainx.ai/blog/seedance-2-5-flash-byteplus-0-152-per-second-claimed-vs-verified-2026) and [Fish Audio's controllable drama audio](https://explainx.ai/blog/fish-audio-drama-3-preview-controllable-tts-natural-language-direction-2026). **VibePrevisBench** scores step 1–2—the part studios historically paid previz artists to do in Maya or Blender before dailies.

The paper's Table 1 (reproduced on arXiv) contrasts prior benchmarks: many cover **text-to-video pixels** or **short VQA**, but few jointly require **multi-shot staging**, **timed actions**, **4D scene state**, and **per-fact ground truth**. VibePrevisBench marks every column yes where competitors are partial or no.

Video storyboard panels illustrating VibePrevisBench graybox proxy videos before photorealistic generation

## Task design: one brief, four input modes

Each task gives an agent:

- A **natural-language brief** (and sometimes images or reference clips).
- A **Blender environment** where it must write and execute Python.

Success means an **importable scene file** plus a **proxy video** that respects duration, fps, resolution, and codec constraints. Tasks span **long, multi-shot** setups—not single-table coding puzzles.

The site groups cases by **scene archetype** (market, classroom, street chase, etc.). Viewers can **re-rank agents per scene**, highlighting that **no single agent wins everywhere**—useful if you route models like [Tencent Hy4's Blender computer-use claims](https://explainx.ai/blog/tencent-hy4-official-release-770b-vision-computer-use-terminal-bench-44-4-weights-planned-2026) versus lighter harnesses.

## Three delivery gates before scoring

The benchmark zeroes tasks that fail infrastructure checks:

1. **Code runs** — valid syntax, standalone execution, tagging toolkit untouched.
2. **Renders complete** — full frame range decodes at specified resolution and fps.
3. **Scene opens** —`.blend` reloads with correct timeline settings.

Only then do **Scene Check** and **Screen Check** run. This design punishes agents that **narrate success** in chat without shipping binaries—a failure mode common in [SWE-GAME-style game builds](https://explainx.ai/blog/swe-game-benchmark-coding-agents-godot-games-opus5-below-60-2026).

### Scene Check (rule-based)

Ground-truth facts become **measurable predicates** on exported scene state: subject positions, routes, turn timing, camera paths, and cut points. Checks use **registered tolerances** and partial credit. Example on the site: start-point error **0.46 m** with subscores for route vs corner placement.

This is closer to **simulation regression testing** than to BLEU on captions—aligned with [envisioning agentic robotics surveys](https://explainx.ai/blog/envisioning-agentic-robotics-survey-s0-s1-s2-roadmap-2026) that separate **world models** from **policy chatter**.

### Screen Check (VLM judge)

A **Gemini 4** judge (with cross-checks against **GPT-6.1 Sol** and **Qwen3.8 Omni Flash**) answers structured questions about the proxy video: recognizable subjects, readable actions, visible key moments, and blind story recap order. Judges score only facts Scene Check already passed, preventing "looks pretty" passes when geometry is wrong.

Authors report **Spearman 0.92–0.96** rank correlation between judges across **2,628** shared runs and **0.90–0.94** correlation with **242** human pairwise preferences—strong for leaderboard stability, though absolute VLM scores differ by ~14 points.

Agent failure modes diagram for coding agents that pass syntax gates but fail VibePrevisBench spatial checks

## Findings: multimodal space beats eloquence

Across **11 agents**, the paper highlights:

1. **Multimodal understanding** separates leaders—especially**3D space** and**camera motion** from flat text briefs.
2. **Basic staging** (standing subjects, simple pans) is often solved;**hard constraints** (precise corners, multi-beat timing) break weaker agents.
3. **Self-inspection loops** exist—agents render previews—but**error correction** still fails, mirroring[Claude Opus TouchSynth PCB](https://explainx.ai/blog/claude-opus-touchsynth-agentic-analog-synthesizer-pcb-2026) stories where vision without verification wastes iterations.

The public site opens **two typical cases per scene type** with all **11 agents'** proxy videos and downloadable **interactive 3D** exports—far richer than a single aggregate accuracy number.

## Seedance, Kling, and builder takeaways

Kling AI's video products compete in the same **director-controlled generation** space as ByteDance's Seedance line. VibePrevisBench makes the division of labor explicit: **agents block**; **diffusion paints**. Teams optimizing inference cost on [Seedance 2.5 Flash](https://explainx.ai/blog/seedance-2-5-flash-byteplus-0-152-per-second-claimed-vs-verified-2026) should still budget **agent GPU time** for Blender iterations—cheap video per second does not fix bad previz.

Practical checklist:

| Step | Action | 
|---|---|
| Baseline | Pick one scene type on the site and rerun your agent harness against published briefs | 
| Metrics | Track **Scene Check** dimension failures, not only final MP4 aesthetics | 
| Harness | Compare **official vendor coding agents** vs generic Terminus-style wrappers | 
| Product | Gate user-facing "script-to-video" on **proxy pass rate** , like[Cinecode](https://explainx.ai/blog/cinecode-claude-code-skill-films-from-code-three-js-ffmpeg-mit-github-trending-2026) gates film skills on ffmpeg outputs | 

## What builders ask after reading the leaderboard

Teams comparing **MCP or skill-based Blender control** ([Google Stitch skills](https://explainx.ai/blog/google-stitch-skills-agent-skills-mcp-claude-code-cursor-install-2026)) against monolithic agent loops want to know whether skills reduce spatial errors. VibePrevisBench suggests **item-by-item verification** against the brief beats a single "render and hope" pass, because Screen Check only awards points when Scene Check facts already hold.

## Judge agreement and why ranks stabilize

The site publishes a **correlation matrix** between VLM judges and humans that matters for anyone borrowing Screen Check as a CI gate. Across **2,628** runs scored by **Gemini 4**, **GPT-6.1 Sol**, and **Qwen3.8 Omni Flash**, **Spearman ρ** between judges runs **0.94–0.96**, and **Pearson r** on agent means reaches **0.97–0.99**. That means a cheaper judge model can substitute for leaderboard **ordering** even when absolute scores shift.

Against humans, the paper reports **242** pairwise duels—one pair of proxy videos per task—with annotators picking the clip that is **easier to understand**. **Spearman ρ** between human win rates and each judge's mean scores lands **0.90–0.94**. Practically, if your agent beats another by **10+ points** on Scene Check, humans usually agree on the winner; if agents tie within noise on geometry, do not trust Screen Check alone to break the tie.

Authors note **GPT-6.1 Sol** declined **20** runs under its content policy and that **14** runs failed rendering gates—those exclusions are why explainx.ai treats the public leaderboard as **research-grade**, not a vendor **SLA** certification. Replicate the gates locally before you paste VibePrevis numbers into a pitch deck.

## Interactive cases and what to watch in replays

The flagship UX on [vibeprevisbench.github.io](https://vibeprevisbench.github.io/?ref=explainx) is not the aggregate bar chart. It is the **per-scene case browser**: two representative tasks per scene type, each showing **inputs**, **eleven proxy videos**, and downloadable **interactive 3D** scene files. When you scrub timelines, look for **recurring failure families**:

- **Camera drift** — agents nail subject placement in Scene Check but VLM Screen Check marks**partial** because heads clip at frame edges during close-ups.
- **Route errors** — pedestrian paths pass coarse distance checks yet miss**turning maneuvers** called out in briefs; the site quotes**0.46 m** start-point error as an exemplar.
- **Multi-shot continuity** — agents stage shot one correctly then**reset world state** between cuts, failing hidden continuity facts even when each isolated render looks fine.

Those patterns mirror [DeltaTau bench](https://explainx.ai/blog/delta-tau-bench-agents-evolving-tool-ecosystems-steward-42-percent-2026) findings that agents handle **single-tool** success but stumble when **environment state** must persist across turns—except here the environment is literally a **timeline**.

For studios already using [Claude Code skills for film pipelines](https://explainx.ai/blog/cinecode-claude-code-skill-films-from-code-three-js-ffmpeg-mit-github-trending-2026), VibePrevisBench suggests adding a **Blender previz skill** whose acceptance tests are **Scene Check dimensions**, not subjective director notes. Pair that with [Seedance 2.5 Flash](https://explainx.ai/blog/seedance-2-5-flash-byteplus-0-152-per-second-claimed-vs-verified-2026) only after proxies pass, so you do not waste diffusion budget beautifying **wrong blocking**.

## How this compares to other agent benchmarks

| Benchmark family | What it measures | VibePrevisBench difference | 
|---|---|---|
| SWE-bench / Terminal-Bench | Patch repos, run tests | Measures **3D scene + video** deliverables | 
| VBench / video quality suites | Final pixel fidelity | Scores **graybox layout** upstream of diffusion | 
| Robotics sim evals | Physics and control | Focuses on **director brief → Blender code** | 
| [SWE-GAME](https://explainx.ai/blog/swe-game-benchmark-coding-agents-godot-games-opus5-below-60-2026) | Game builds in Godot | Same **long-horizon** vibe, different engine | 

None of those replace VibePrevisBench; they explain **where to plug it** in a CI matrix. Use Terminal-Bench for **tooling**, VibePrevisBench for **previz**, and Seedance for **final pixels**—three gates, three budgets.

## Limitations

- **November 2026 release** — confirm task JSON and harness drops on arXiv updates before baking into CI.
- **Judge dependence** — Screen Check inherits VLM blind spots; human agreement is high on**ranks** , not every edge frame.
- **Kling affiliation** — authors include Kling AI; treat leaderboard ordering as**research** , not a vendor certification.
- **Compute** — full 242-task sweeps across 11 agents are not laptop-friendly; sample by scene type first.

## Related reading

*Paper version, site copy, and model names are accurate as of arXiv 2611.15789 and vibeprevisbench.github.io accessed November 16, 2026.*
