Introduction

Most image-to-3D tools want a clean single view or a posed capture. UniForge3D takes the messier input game teams actually have: a handful of unposed photos around an object. From Peking University, Math Magic, and HKUST, the drop ships Apache-2.0 code plus live Hugging Face weights — not a paper teaser. Clone, run setup.sh, drop PNGs, fire infer.py.

What shipped

UniForge3D jointly runs reconstruction and generation in a shared anchor-guided space. Reconstruction locks the geometry you already photographed. Generation fills the missing sides with learned shape priors. The shared space follows the anchor view’s azimuth while keeping the object upright, so the two branches can exchange features without fighting camera-relative vs canonical coordinates.

Structure-guided view selection then picks the most informative input view per voxel and feeds those image features into detailed shape generation. The team reports state-of-the-art appearance and geometry on Toys4K and GSO. Stack notes: the repo builds on TRELLIS.2 and VGGT.

Primary surfaces:

Paper badge is still “Coming Soon.” Treat the claims as vendor/author benchmarks until the arXiv PDF lands.

Pro Tip

For prop and pickup pipelines, shoot 3–6 casual angles around the object — no COLMAP, no pose files. UniForge3D expects a per-case folder of PNGs under assets/. That is closer to a field capture than a studio turntable.

Run it today

git clone https://github.com/xibushu/UniForge3D.git
cd UniForge3D
bash setup.sh
conda activate uniforge3d

Layout multi-view inputs like this:

assets/
├── case01/
│   ├── 001.png
│   ├── 002.png
│   ├── 003.png
│   └── 004.png
└── case02/
    └── ...

Then:

python infer.py

Weights live on the HF card under uniforge3d/ (stage1 joint structure model, shape512/shape1024 denoisers + adapters, texture1024, VGGT backbone/conditioner). This is local GPU inference — no hosted playground yet. Closed-API alternatives (Meshy, Tripo, Scenario) still win if you need one-click Studio exports; UniForge3D wins if you want open weights and multi-view completion under your own roof.

Why it matters for game teams

Building Rome turned a single photo into a metric scene mesh. Thrixel pushed named parts for agents and engines. UniForge3D sits in the same asset-gen lane with a different bet: complete an object from sparse casual views instead of hallucinating from one hero shot. For environment props, collectibles, and reference-based kitbashing, that is the practical capture pattern.

Expect rough edges on day one — low community traction, paper still pending, local setup only. The runnable surface is real: code, license, and weights are public now.

Conclusion

If your backlog is “we have four phone photos of this prop,” UniForge3D is worth a same-day clone. Open weights, clear infer path, joint recon+gen with view selection — and a paper to watch when it drops.

—Titus