Skip to main content
Back to timeline
openbiox BizardSource publication:

Hi3D v1.0 goes open source: a reference image turns a coarse mesh into a 1536³ sparse extraction grid

Synopsis

Hi3D v1.0, released on October 23, 2026, ships inference code and eight model files that refine an existing coarse 3D mesh into higher-resolution geometry conditioned on a reference image of the same object. It builds on the Sparc3D sparse representation and refinement architecture, progressively sampling sparse latents, decoding them with a self-pruning sparse VAE and extracting a mesh; the project reports training on roughly two million curated 3D assets and refinement up to an effective sparse extraction-grid resolution of 1536³. The repository contains inference only, so the training code, evaluation harness and comparison meshes are absent and the report's comparative tables cannot be reproduced from it alone.

AI-generated editorial illustration: 「牛来」将彻底成为过去式!Hi3D重塑三维模型新高度,V1.0正式开源

Interpretation

The 3D reconstruction pipeline gains a refinement step that can be called on its own: it takes a reference image of the same object plus a coarse mesh and returns a geometrically finer mesh, generating no textures, materials or vertex colors. Previously coarse generation and refinement came as one bundle; the Hi3D v1.0 refinement script has no Hunyuan dependency, and the coarse mesh can be supplied by the user or through a replaceable adapter. The repository's comparisons are a local-geometry figure across nine methods and multi-view figures for four objects, with Hi3D v1.0 highlighted in teal in the rightmost column and rendered as shaded geometry and surface normals.

Refinement runs in passes by resolution, and each pass revoxelizes the mesh extracted by the previous one to provide the sparse support for the next. The three passes pair nominal structural grids of 64³, 128³ and 192³ with sparse extraction grids of 512³, 1024³ and 1536³; a 512³ request runs one pass, 1024³ runs two and 1536³ runs all three. Image features condition sparse flow on coordinates derived from the coarse geometry, and decoding and mesh extraction are followed by revoxelization; these labels describe nominal sparse coordinate and extraction grids rather than a dense 1536³ latent tensor or a fixed triangle count.

Perspective

The demonstrated scope is geometry refinement of a single object at batch size one: a reference image and a coarse mesh go in, and the output mesh is recentered and normalized to unit maximum radius, so world coordinates and absolute scale are not preserved. The 1536³ ladder has only been validated on a 96 GB NVIDIA H20, and peak memory and feasibility on smaller cards have not been established; the repository provides inference code only, without training code, evaluation harness, benchmark splits or comparison meshes.

Severe structural errors in the coarse mesh can persist, and geometry hidden from the image is a plausible completion rather than an exact reconstruction. GPU sparse convolution and attention are not bitwise deterministic, so seeds are not mesh hashes and meshes should be compared geometrically under the same inputs, artifacts, scheduler and environment. Adaptive voxelization may use a coarser support grid on complex inputs and emits a warning, and a request that cannot reach the target resolution raises instead of silently returning a lower one.