OctLLM: Octrees as an Explicit 3D Language

Demo of OctLLM (code, weights).

• Image → 3D and Text → 3D: OctLLM generates a Sparse-Octree (S-Octree) token sequence, a 3D U-Net completes the occupancy, and TRELLIS decodes a textured GLB.
• 3D understanding: an uploaded GLB/OBJ is converted into an S-Octree sequence and OctLLM answers questions about it.

Inference code (llamafactory/, trellis/, scripts/, dataset_toolkits/, inference.py) is vendored from the official repository (MIT for original OctLLM contributions; third-party notices in LICENSE). Example images and meshes are from the official repository's assets/examples.