UniForge 3D turns a few sparse photos into complete 3D models

UniForge 3D is a newly open-sourced tool that turns a handful of casual photos into a complete 3D model, filling in every blind spot you missed with the camera.

Traditional photogrammetry usually falls apart when parts of an object are occluded or unphotographed, leaving holes and broken meshes. UniForge 3D takes a small set of unposed, sparse images and reconstructs both the visible parts and the missing backside geometry.

Reconstructing what is seen, imagining what is not

The trick lies in merging direct geometric reconstruction with generative 3D shape priors into a single pipeline. Reconstruction locks down every visible detail from your input shots, while the generative side predicts what unseen regions should look like.

Previous attempts often struggled because reconstruction happens relative to the camera while generative models prefer a normalized canonical space. UniForge 3D solves this coordinate mismatch by aligning everything in a shared anchor-guided space, keeping the object upright while following the reference viewpoint.

The system exchanges features bidirectionally between point-map predictions and a complete sparse voxel structure. Crucially, it links each 3D voxel back to the most relevant input view rather than blending all angles together, which prevents fine surface details from being blurred out.

What it handles and how to run it

According to the authors, the model performs especially well on character figurines, sculptures, and complex household objects. Benchmark tests on Toys4K and GSO show state-of-the-art results in both geometric fidelity and visual appearance compared to existing methods.

The research team has open-sourced both the code and model weights. You can grab them from the project repository, where instructions are provided for setting up the environment and running inference locally on your own machine.

Why it matters

For indie game devs, digital sculptors, and 3D printing tinkerers, this is an easy recommendation to test out. It cuts down the grueling requirement of shooting dozens of overlapping photos with dense camera calibration to just a few casual snaps.

Pure 3D generative models often hallucinate plausible shapes that ignore your exact input, while classic scanning tools leave gaping holes. UniForge 3D hits a smart sweet spot by letting high-precision reconstruction guard the real pixels while letting generation cleanly patch the gaps.