A new image-to-3D system claims it can finally stop AI from mangling the logos and text printed on the objects it reconstructs.
Hi3D 3.0 is an image-to-3D generation system built around a geometry model called Twinkle3D, which produces watertight 3D meshes at 2048-cubed resolution. The paper targets a specific failure: inscriptions, brand marks, and repeated patterns that usually get smeared or dropped when a photo becomes a 3D mesh. The team fixed surface-quality and watertightness problems in its prior geometry representation, scaled the diffusion model to handle sequences of up to 300,000 geometric tokens, and cut training time per step from about ten minutes to about ten seconds via a redesigned architecture and distributed training. They also front-loaded shape and detail work into the first generation pass, skipping the two-stage refinement pipeline most rivals rely on.
The numbers are the real story: Hi3D 3.0 recovered 82.1% of inscribed characters at 98.2% precision, against 21.7% recall for the best of four commercial systems tested. That gap matters because generic 3D generators optimize for objects that look plausible, not objects that are provably the one in the photo - fine for a stock chair, useless for a product shot where the label is the whole point.
Worth noting: the comparison is against four unnamed commercial tools on the paper's own benchmark, not an independent leaderboard, so treat the margin as a claim to watch rather than a settled result.