Inception3D

Imagination, grounded in a world model

Spatial intelligence remains passive if it can only reproduce what has already been seen. Synthesis must handle sparse observations, missing views, and dynamics—without drifting away from geometric constraints.

Closing the gap between reconstruction and generation

Reconstruction usually needs dense observations, whereas generation often begins with one view—or none at all. GenFusion conditions video diffusion on artifact-prone RGB-D renderings, then cycles restored frames back into reconstruction training. Generation no longer ignores geometry: it repairs the missing evidence that geometry makes explicit. Reconstruction, in turn, expands through new plausible views. Representation constrains synthesis; synthesis expands representation.

GaVS brings the same principle to video stabilization, reframing it as temporally consistent local 3D reconstruction and rendering. Scene extrapolation avoids the cropping typical of 2D stabilization, so each frame is rendered again within an interpretable local world.

From static shape to 4D worlds

A dynamic world concerns both what exists and how it changes. Motion 3-to-4 decomposes 4D synthesis into static 3D shape and motion reconstruction. A canonical mesh, a compact motion latent, and per-frame vertex trajectories recover complete, temporally coherent geometry from monocular video and limited 4D data. The decomposition turns one difficult problem into reusable structure for content and change.

From reconstructed motion to editable objects

Recovering motion is only the first step; dynamic objects must be controllable for creation, simulation, and interaction. GaussiAnimate extracts and refines skeletons from dynamic Gaussian or mesh sequences, binding motion with skelebones that capture both kinematic structure and non-rigid surface dynamics. Humans, clothing, quadrupeds, and other deformable categories can then be re-rigged and reanimated without discarding rich surface detail.