Peking University's UltraShape 1.0 enables high-fidelity 3D generation via geometric refinement
tom_doerr · x · 2026-08-15
Peking University's YuanGroup released UltraShape 1.0, a scalable diffusion framework for high-fidelity 3D shape generation. The method uses a two-stage pipeline: first synthesizing a coarse global structure, then refining it via voxel-based techniques to produce detailed geometry. The project also introduces a comprehensive data processing pipeline featuring novel watertight processing and high-quality filtering to enhance public 3D datasets by removing defects and preserving fine details.
More from Multimodal
- Testing if AI can recreate Mr Bean style — Ok_Roll_8698 · 2026-08-15
- Custom nodes for Minimax latent tools enable video upscaling and detail enhancement — Striking-Long-2960 · 2026-08-15
- Director: AI video creation sphere is most inspiring in 15 years — LudovicCreator · 2026-08-15
- Pika releases 4 audio models, cuts costs by 20x — dr_cintas · 2026-08-15
- MiniMax H3 vs LTX 2.5: Controllable camera paths and timeline in ComfyUI — RobbaW · 2026-08-15
- AI recreation of Pulp Fiction scene looks shockingly real — taherdhanera · 2026-08-15