Building a complete 3D asset from a few casually captured images is difficult because the input is sparse and lacks camera poses. A new arXiv preprint, UniForge3D, frames this as a joint reconstruction-and-generation problem: reconstruction should preserve the geometry that is visible, while generation should plausibly complete the rest.
The paper describes a shared anchor-guided space that links these two objectives. By anchoring both reconstruction and generation in a common representation, the method is designed to generate full 3D assets without dropping the details that were observed in the input images.
As a single preprint, the work has not yet been peer-reviewed. Its stated contribution is a unified space where both tasks can be trained and applied together, rather than treating reconstruction and generation as separate steps.