metadata
pipeline_tag: image-to-3d
FreeInpaint
Pose-Free Feed-Forward 3D Inpainting via Learnable Mask Attention and Support Token Refinement
Presented at NeurIPS 2026.
- Paper: arXiv
- Project page: https://rorisis.github.io/FreeInpaint
- Code: https://github.com/Rorisis/FreeInpaint
FreeInpaint is a feed-forward framework that generates complete and 3D-consistent scenes directly from unposed multi-view images with masked regions, eliminating the reliance on pre-computed camera poses. It introduces a Learnable Mask Attention mechanism and a Support Token Refinement strategy to achieve high-quality inpainting while preserving geometric and textural consistency.