AI video generators are good at making things look real and poor at following exact directions: where the car turns, how high the camera sits, when the door opens. One fix is to give them a rough 3D animation first, a grey-box proxy that fixes layout, motion and camera, and let the model paint the realistic detail on top. VibePrevisBench, a preprint from Kling AI with Tianjin University and other partners, tests how well coding agents can build those proxies in Blender.
Each of its 242 tasks gives an agent a brief, which may be a script, a set of photos, the first frame of a scene to continue, or a reference video. The agent writes Blender code to build a 4D scene, a 3D scene that changes over time, and renders it. Scoring has two parts. A rule-based check re-runs the code and compares the exported scene with hidden ground-truth facts in nine dimensions. A vision-language model then judges whether the facts that passed can actually be seen in the video.
GPT-6 Astra led the 11 agents with 77.7, followed by GPT-6.1 Sol at 76.1, Gemini 4 Argon at 75.0 and Claude Opus 5.5 at 74.1; Muse Spark 1.3 was last at 61.1. Busy street scenes were the hardest domain for every agent, and storytelling the weakest visual dimension. The authors' three findings: visual understanding of space and camera movement separates the agents; all of them handle basic camera work and staging but fail on hard requests; and agents look at their own renders yet rarely fix what is wrong.
Cost varies widely. Astra averaged $32.59 per task, against $7.82 for Opus 5.5, $6.71 for GPT-6.1 Sol and $4.45 for Gemini 4 Argon. More turns went with lower scores. For anyone driving Blender through an agent, for example with MCP for Blender, the practical advice is to make the agent check its render item by item against the brief.
The limits: this is a preprint from a video company with an interest in the result, and the visual judge is itself a model, with its own blind spots. Two examples show proxy errors carrying into the final video, but how often that happens was not measured.