A new technical report on arXiv, dated October 8, 2026, lays out Hi3D 3.0, an image-to-3D generation system built around a geometry model called Twinkle3D. Hi3D 3.0 recovers 82.1% of inscribed characters at 98.2% precision; the best of four rivals it was tested against managed 21.7% recall.
Twinkle3D generates watertight triangle meshes at 2048³ resolution. The authors say they fixed the surface-quality and watertightness problems that an earlier approach (O-Voxel/FaithC) had at that same precision level, scaled the diffusion generation to sequences of up to 300,000 geometric tokens through a redesigned DiT architecture, and cut training time per step from roughly ten minutes to ten seconds, the authors say, through distributed training changes. They also moved away from a two-stage generate-then-refine pipeline, arguing that refinement cannot fully fix errors baked in at the first stage, and claim their single-stage model beats prior two-stage pipelines that used 512³ refinement.
The paper reports Hi3D 3.0 was evaluated against Hi3D 2.1, Meshy 7, Tripo 3.1, and Rodin Gen 2.5 on silhouette alignment, normal-field detail alignment, multi-view completeness, shape aesthetics, and inscription readback, and says it outperforms all four across every reported metric. The report does not state sample size for the inscription test, pricing, release availability, or hardware requirements for running Twinkle3D.
Hi3D 3.0’s own inscription test doesn’t disclose how many objects were used, the same gap that undercut Meshy’s 97% claim when 3Dnatives checked it: that review found the figure rested on just 15 models.