UniForge3D Builds Complete 3D Models From a Few Unposed Photos
A new sparse-view 3D asset pipeline fuses reconstruction and generation in a shared coordinate space, with code and weights now public.
- UniForge3D jointly runs 3D reconstruction and generation inside one shared anchor-guided coordinate space.
- Anchor view defines azimuth, keeping objects upright and eliminating camera-relative vs canonical mismatch.
- Each voxel selects its most informative input view instead of averaging features across all views.
- Reports state-of-the-art appearance and geometry on Toys4K and GSO sparse-view benchmarks.
- Code and model weights available on Hugging Face and the project page.
- Targets object-level assets from unposed sparse photos, not full scene reconstruction or PBR materials.
UniForge3D is a new system for generating complete, textured 3D meshes from a small set of unposed object photos. The model combines geometric reconstruction with learned shape generation inside one architecture, allowing it to preserve visible details while filling surfaces that no input camera captured. The paper reports state-of-the-art geometry and appearance results on the Toys4K and Google Scanned Objects benchmarks.
Sparse-view reconstruction remains difficult because each available technique solves only part of the task. Reconstruction models recover geometry supported by the images but leave gaps around occluded surfaces. Generative models can infer those missing regions, although their predictions may drift from the observed shape. UniForge3D coordinates both processes through a shared reference frame and voxel-level feature routing.
| Component | UniForge3D |
|---|---|
| Input | Sparse, unposed images of one object |
| Camera metadata | No ground-truth poses required |
| Output | A complete 3D mesh with predicted appearance |
| Core technique | Joint reconstruction and generation with voxel-wise view selection |
One frame for two jobs
Reconstruction networks typically predict geometry in camera-relative coordinates because pixels describe an object from a particular viewpoint. Generative 3D models usually learn shape priors in a canonical object space, such as an upright, front-facing orientation. Features expressed in these different coordinate systems cannot interact reliably without an alignment step.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
