GraphWrit3R replaces multi-stage 3D scene-graph pipelines with an end-to-end model that emits objects, attributes, and relationships as structured JSON. It accepts point clouds, Gaussian Splats, or both, and supports open-vocabulary queries through its language-model decoder.
Point clouds and Gaussian Splats are encoded separately, projected onto a shared voxel grid, and aligned with a per-voxel contrastive loss before decoding. On 3DSSG, the system reports state-of-the-art object-class, predicate, and triplet recall, including against methods given ground-truth object annotations at inference. This is interesting for deployable 3D understanding because it removes explicit intermediate representations and proprietary model calls while supporting multiple 3D input formats with one set of weights.