BVB Benchmark: Assessing Video Reconstruction in Blender with AI Agents

More and more videos are appearing online where large language models reconstruct interiors in Blender format. This process can be called LLM photogrammetry.

The BVB benchmark tests 51 agent configurations on 288 real video clips of interiors. The test takes place in a unified environment with the same prompt and a strict budget limit.

Limitations and Evaluation Metrics

Agents are prohibited from using ready-made 3D assets from external libraries. All geometry, materials, and animations must be created programmatically via primitives and operators within Blender itself.

Each scene is evaluated against 5,130 spatiotemporal questions. Assessment includes preservation of original information and visual similarity during rendering. Failed reconstructions receive a zero score.

Model Testing Results

• Fablanski took first place with a score of 5.2
• Astrachka came in second with a result of 6
• GPT Sol secured third place with a score of 6.1
• Argon agent ranked fifth
• Opusian achieved 5.5 (sixth place)

The surprise was the tenth-place finish by the cheapest model, Inkling 2 Small from Mira Murati. All DeepSeek models with Kwenas ended up in the top ten. Information on Kimi's results is unavailable.