🤖 NVIDIA introduces ArtiFixer to improve 3D reconstructions
The Wan2.1-based model (16.9 billion parameters) uses a two-stage pipeline to generate hundreds of frames that fill gaps in 3D scenes.
🌍 ArtiFixer solves the problem of poor extrapolation in areas with insufficient coverage, which is critical for methods like 3D Gaussian Splatting. This enables the scaling of new viewpoint generation without losing consistency.
👤 This represents a qualitative leap in creating detailed 3D worlds from a limited set of photos. Now, "holes" in scenes are filled with realistic content that strictly follows the object's structure.
Source 1: https://research.nvidia.com/labs/sil/projects/artifixer/ Source 2: https://github.com/nv-tlabs/artifixer