AuroraRyan/PartTrellis-corpus

PartTrellis corpus — 4-source part-labelled training data

The full training and test corpus of PartTrellis: 19,132 training objects (lists/tri_train_dual.txt) and 1,000 held-out test objects (lists/test1000.txt), drawn uniformly from four part-labelled sources: Articraft, PartNeXt, Fusion360, and TRELLIS-500K assets. Includes the original textured meshes, the rendered conditioning views, and all precomputed training targets (O-Voxel dual volumes and latents), so both training and evaluation can run without re-doing preprocessing.

Layout

shards/assets_000.tar …          # ~1,000 objects each (train ∪ test)
  <source>/<oid>/mesh.glb            original textured mesh
  <source>/<oid>/face_ids.npy        per-face part labels
  <source>/<oid>/meta.json
  cond/<source>/<oid>/renders_cond_pbr/   PBR conditioning views + transforms.json

latents/latents_000.tar …        # 5,000 training objects each
  <source>/<oid>/A.vxz  B.vxz        packed dual O-Voxel volumes (Stage-1 target)
  <source>/<oid>/tri_ss_latent.npz   sparse-structure latent
  <source>/<oid>/tri_slat.npz        SLat latent (legacy fitted-VAE space)
  <source>/<oid>/tri_meta.json       bipartite contraction metadata
latents/origslat.tar             # SLat latents in the ORIGINAL TRELLIS.2 VAE
                                 # space — the Stage-2 target of the released model

eval/gt_cache.tar                # cached GT part point samples for the evaluator
lists/                           # train/test/per-source curation lists + recipes

Note on Stage-2 latents: the released PartTrellis Stage-2 (stage2_slat_orig_100k.pt) trains on latents/origslat.tar (original TRELLIS.2 VAE space). tri_slat.npz inside the latent shards is the legacy fitted-VAE space kept for reproducing older ablations; do not mix the two.

Preprocessing provenance

Meshes are voxelized as-is after rigid normalization only (centering, unit cube, Z-up) — O-Voxel is a field-free surface proxy, so no watertighting, remeshing, or cleaning is applied. Part contact graphs are contracted until bipartite and packed into two volumes (A.vxz/B.vxz). All preprocessing scripts ship in the code tarball of https://huggingface.co/AuroraRyan/PartTrellis (attraction/data/).

Licensing

The packaging and lists are released for research use; every asset remains under the license of its original distribution.