LEGO-Anything: coding agents write Blender code to rebuild editable 3D scenes, 62.7% gains
AWS · hf · 2026-09-30
LEGO-Anything is an Image-to-Code framework where a coding agent iteratively writes and executes Blender code, inspects renderings, and revises an explicit scene program instead of emitting a fixed 3D output.
- Ships with LEGO-Bench: 208 images from 104 indoor/outdoor scenes, scoring artifact validity, visible-surface geometry, and rendered appearance separately
- GPT-6-astra scores best overall (53.4% indoor, 39.6% outdoor), yet large gaps remain between producing valid artifacts and faithful geometry/appearance recovery
- Trajectory analysis reveals recurring failure modes: weak scene initialization, regressive edits, unreliable self-evaluation
- LEGO-Plugin, a training-free harness plugin, improves all six evaluated agents with up to 62.7% relative overall gains
- LEGO-World shows reconstructed scenes can derive detections, masks, and depth as queries, though well short of specialized vision models
More from coding & agent
- Auto-research loop on 120 B300s finds 40% Kimi K3 inference gain for $9,176 — bookwormengr · 2026-09-30
- 8 research agents, one 30B model, 144 hours: RSIArena tests AI-driven post-training research — my_cat_can_code · 2026-09-30
- LENINROOMS: Building an Infinite Soviet Apartment Backrooms With Claude — teortaxesTex · 2026-09-30
- dart-query MCP server adds DartQL bulk task operations to cut token usage — modelcontextprotocol · 2026-09-30
- ISIR MCP connector lets agents query Czech insolvency register by company or person — modelcontextprotocol · 2026-09-30
- New Yorker-style illustration Skill hits 400 stars, monetizes via Baidu agent ecosystem — oran_ge · 2026-09-30