BDXXN/Finetune_neurok

Finetune NeuROK deformation corpus

This repository stores a standalone, NeuROK-compatible deformation corpus under objaverse/finetune_deformation. It combines processed public animation data with generated MPM solid trajectories. Training consumers should read index/all.jsonl (or its train.jsonl and val.jsonl split files) rather than discovering NPZ files by directory traversal.

Corpus composition

The validated public-animation branch contains:

• —1,998 DynamicObjaverseProcessed geometries;
• —500 DyMesh / AnimateAnyMesh geometries;
• —2,498 trainable rows in the canonical combined index;

The generated target is 5,000 geometries × 2 accepted clips = 10,000 clips. Its deterministic 24-cell cycle balances elastic, plastic, sand, and snow, with every material parameter drawn from one of four strata before a stratum repeats. Later rounds append previously unseen cells without changing prior samples. The active external conditions are drop, press, jet pair, shear, and uniform gust. Drop cycles through upright, 85–95° quarter-turn, and random orientations at least 30° from upright; every variant starts from rest and retains the single drop condition label.

Generated geometry retains every modeled object component, including plates, fruit, cream, and decorations. Only recognized reference helpers such as source-scene ground sheets and the exact Blender bone-display Icosphere fingerprint are filtered. Retained components are simplified independently and voxelized into one watertight MPM simulation body while the visible surface keeps its original fixed component topology.

The preparation grid, MPM particles, stored surface, and training points are distinct. A 128-cell voxel grid is used once to repair the simulation solid. Generation then uses a 160 cells/m grid and starts every geometry at 4 mm MPM particle spacing. A geometry that produces at most 32,768 particles is kept exactly at 4 mm; only an oversized geometry is deterministically resampled at a coarser spacing until its actual PBS particle count is at most 32,768. This check happens before the MPM grid is allocated. Accepted clips store the 32-frame fixed-topology visible surface; NeuROK training samples 512 corresponding surface points from each rest/target pair rather than feeding every stored mesh vertex to the network.

All sources use the same conservative global SE(3) decision: global pose is removed only at high confidence, leaving the model responsible for deformation. An accepted generated clip must have strictly more than 10% bbox-normalized peak deformation after this handling; press must also achieve strictly more than 20% height compression. Nonfinite state, lateral/upper numerical-domain contact, excessive deformation, invalid bonded-material topology, or excessive granular spread is rejected. Floor contact is allowed. Each planned cell is drawn once and simulated once; a failure is replaced only by a previously unseen cell in a later diversity round, never by per-cell search or retry.

Layout

objaverse/finetune_deformation/
  dynamic_objaverse/
  dymesh_16f/
  dymesh_32f/
  generated_mpm/
  index/
    all.jsonl
    train.jsonl
    val.jsonl
    summary.json
  schema.json

Source code and the complete machine handoff are maintained at https://github.com/AE-W/phsam4d-finetune-dataset.