Point2Part Redefines 3D Part Decomposition as Full Partitioning, Outrunning GPT-6-Astra 559-to-1
Scobleizer · x · 2026-11-03
A CMU-led team introduced Point2Part, a fast feed-forward unified model that reframes 3D part decomposition as a complete, non-overlapping partition of the whole shape. Users interactively specify parts with one 3D point each.
- One model handles image→parts, mesh→parts, and mesh segmentation, achieving SOTA part quality with >10× lower inter-part penetration
- Benchmark against GPT-6-Astra (released the same month): comparable quality, but Astra spent 514K tokens ($1.25) on one object while Point2Part finished 559 objects for free in the same time
Paper and project page are public.
More from Multimodal
- Runway launches Runway Dev, an API platform built for coding agents — runwayml · 2026-11-03
- MOSS-Image ecosystem roundup: smart sampler, handwriting LoRAs, 7.5GB WebGPU build — optimisticalish · 2026-11-03
- Single-shot 2-minute drive video with MiniMax H3/Slopus, start frame as only reference — BittiAI · 2026-11-03
- Creator shares full timestamped prompt for photoreal southern gothic swamp creature AI video — LudovicCreator · 2026-11-03
- Industrial designer uses fal to iterate real hardware designs from photos, faster prototyping — OdinLovis · 2026-11-03
- Dev uses Claude Opus 5.5 to build a stunning nanite/voxel-inspired forest — dreamwieber · 2026-11-03