Breaking

Peking University Team Releases 3D Generator UltraShape 1.0 With Two-Stage Refinement

July 2, 2026 at 18:46 EDT

Media Generation
Open Source
Research & Papers

Peking University Yuan Group (PKU-YuanGroup) has released UltraShape 1.0, an open-source framework that generates high-fidelity 3D meshes from images or text. By separating coarse shape generation from detailed refinement into a two-stage geometric process, it aims to approach the quality of commercial systems using public data and limited compute.

December 2025 · PKU-YuanGroup

UltraShape 1.0: Turning Blobby 3D Into Production-Ready Meshes

An open-sourced two-stage framework that generates high-fidelity 3D shapes from a single image — refining a coarse mesh with position-explicit voxel detail to hit commercial-level geometric quality using only public data and limited compute.

Stages: coarse structure then voxel refinement

50→12

Sampling steps, default down to a fast minimum

8192

Latents in Low VRAM mode for smaller GPUs

The Pipeline

Single Image

One input photo

→

Stage 1 · Coarse

Global structure via Hunyuan3D-2.1

→

Stage 2 · Refine

Voxel queries + RoPE anchoring on a DiT backbone

→

.glb / .obj

Production-ready mesh

Geometric Detail Retained

Where existing image-to-3D models go "blobby," refinement preserves fine features.

Coarse only

blobby, soft edges

+ UltraShape

thin parts, holes, sharp edges

Where it wins

Suppresses blobby artifacts for clean meshes
Retains engraved detail and sharp edges
Structural accuracy suited to CAD, simulation, printing
Easy to combine with other coarse-generation tools

Open questions

Needs a separate tool for coarse mesh generation
Tuning required under Low VRAM settings
Complex data preparation despite released code
Large-scale evaluation still accumulating

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.