UniForge3D studies how to produce a complete 3D asset from a few unposed images. Its October 11 paper combines reconstruction of observed geometry with learned priors for regions absent from the photographs.
An anchor-guided coordinate space connects the two branches. The system then selects relevant views for individual spatial elements to preserve details, rather than treating every photograph as equally informative.
The authors provide code and model resources. Their original code is MIT-licensed, but VGGT-derived weights remain under CC BY-NC 4.0, and the default background-removal component has separate commercial terms. Code availability alone therefore does not establish unrestricted commercial deployment.
Sources: Research paper · Author implementation and license notes