Inception3D

Seeing is not yet understanding

An image is not a world. It hides occluded structure, breaks temporal consistency, and entangles objects with lighting, viewpoint, and motion. We seek representations that retain what matters while remaining continuously updatable.

Reading 3D understanding from visual features

Vision foundation models excel at 2D tasks, but depth or matching scores alone cannot tell us whether they understand 3D. Feat2GS reads foundation-model features into 3D Gaussian attributes and separates geometry from texture, making novel-view synthesis both a reconstruction tool and a probe of spatial understanding. LoftUp brings high-resolution images, coordinates, and low-resolution features together through coordinate-based cross-attention, returning semantic evidence to sharp pixel boundaries.

Disentangling the visible world

A representation that renders well is not necessarily easy to reason about or edit. Ref-GS separates direction, roughness, geometry, and lighting to reduce the ambiguity between surface normal and viewing direction. NeST places texture in a neural shell around the surface, freeing Gaussians to focus on geometry and sampling. It preserves detail with fewer primitives and supports direct textured-mesh extraction. A useful representation should not only look right; it should place distinct sources of variation into variables that can be analyzed and controlled independently.

From one-off reconstruction to online world memory

The real world does not appear all at once. Easi3R separates camera and object motion from the attention of a pretrained model without additional training; Human3R jointly recovers people, scene, and camera trajectories in a shared world frame. For long video streams, TTT3R updates memory according to confidence in state-observation alignment, while R³ uses confidence-weighted relative constraints for offline and bounded-memory streaming reconstruction. GlobalSplat aggregates multi-view evidence into global scene tokens before decoding explicit geometry, bridging compact latent state and usable 3D structure.