BVB Benchmark Tests Agent Video Understanding by Rebuilding 288 Real Videos in Blender
SonglinYang4 · x · 2026-10-14
Researchers from Rochester, Sony, CMU and UW introduce BVB, built on the idea that 'if an agent truly understands a video, it can reconstruct it programmatically.'
- Agents watch real indoor videos and write Blender Python code to reconstruct them as animated 3D scenes with a camera following the original trajectory, requiring understanding of layout, objects, and viewpoint changes over time.
- 288 real videos, 51 agent configurations; identical conditions: frames only, same sandbox, per-scene cost cap.
- Evaluation on two axes: Dual VQA (semantic preservation) and Latent Similarity.
- Leaderboard now includes Gemini 4 Argon, Muse Wonder, Claude Haiku 5.5, DeepSeek V4.1 Flash, and Grok 4.7, with visible progress in programmatic video reconstruction.
Paper, code, demos and leaderboard are public.
More from Multimodal
- Dev Builds Planetary Lander Sim in Unity Using AI and VR, Says Opus 5.5 Now Outputs Video — ZeroStateReflex · 2026-10-14
- Claude Motion & Dashboards cheat sheet: $8.7K explainers and analyst dashboards for $20/month — rubenhassid · 2026-10-14
- Striking cyberpunk AI art of neon-haired girl against sticky-note wall — LudovicCreator · 2026-10-14
- Dad uses Gemini Argon to generate custom marble run videos for his daughters — jacalulu · 2026-10-14
- Grok Bot chains Suno, Midjourney and Opus to autonomously make music videos — tetsuoai · 2026-10-14
- AI-generated sibling short goes viral: "only I get to pick on my little brother" — alifcoder · 2026-10-14