Pose-Free Feed-Forward 3D Inpainting via Learnable Mask Attention and Support Token Refinement
FreeInpaint introduces a new method to generate complete 3D scenes directly from unposed multi-view images with missing parts, removing the need for precomputed camera poses. It achieves this by using Learnable Mask Attention to handle corrupted cross-view reasoning and a Support Token Refinement strategy to fill in occluded areas using diffusion models, resulting in high-quality, fast 3D inpainting.