Details
- NVIDIA AI announces Axolotl3D, a multimodal 3D generation and shape completion model unveiled at ECCV 2026.
- The system combines images, camera parameters and partial 3D geometry such as point clouds to infer missing object regions.
- Axolotl3D is explicitly occlusion-aware, reconstructing unobserved parts of objects while preserving the geometry that is already observed.
- The model jointly conditions on images, visibility masks, camera data and partial points, using the partial geometry as a stable anchor for faithful reconstruction.
- A unified training framework leverages large-scale 3D datasets to handle single-view, multi-view, occluded and editing scenarios within one model.
- Research experiments on benchmarks like Toys4K and OmniObject3D report state-of-the-art performance in both clean and occluded settings.
- The work targets practical uses such as multi-view image-to-3D from casual photos, occlusion completion and geometry-consistent object editing.
Impact
By consolidating multi-view reconstruction, occlusion completion and geometry-aware editing into a single framework, Axolotl3D strengthens NVIDIA’s position in 3D vision and generative graphics. The model appears to narrow the gap with other research efforts in occlusion-aware 3D generation and could make high-fidelity 3D asset creation more accessible for robotics, simulation, gaming and digital twins.
