Inference Results HTML View: Here

Based on the RigNet dataset (2703 preprocessed 3D character models with ground-truth rigs).

---

Pipeline Overview

Phase	Name	Status
1	Point Cloud Dataset Generation	docs/phase_1.md
2	Shape & Skeleton Tokenizers	docs/phase_2.md
3	Hybrid Attention Transformer	docs/phase_3.md
4	Joint Diffusion Module	docs/phase_4.md
5	Connectivity Prediction	docs/phase_5.md
6	Skinnign Weight Prediction	docs/phase_6.md
7	End-to-End Training	Not started
8	Evaluation and Metrics	Not started

---

Project Structure

riganything-mini/
├── Dataset/                        # RigNet preprocessed dataset
│   ├── obj_remesh/                 # Remeshed OBJ meshes (1K–5K verts)
│   ├── rig_info_remesh/            # Rig info txt files (joints, hierarchy, skinning)
│   ├── pretrain_attention/         # Pre-computed attention supervision (Phase 3/4)
│   ├── volumetric_geodesic/        # Pre-computed geodesic distances (Phase 5)
│   ├── vox/                        # Voxelized models
│   ├── train_final.txt
│   ├── val_final.txt
│   └── test_final.txt
├── pointClouds/
│   └── obj_remesh/                 # Phase 1 output
│       ├── <id>_pointcloud.npy     [1024, 6]  xyz + normals
│       ├── <id>_points.npy         [1024, 3]  positions
│       ├── <id>_normals.npy        [1024, 3]  outward normals
│       ├── <id>_skeleton.npy       [K, 4]     joints + BFS parent index
│       └── <id>_skinning.npy       [V, K]     skinning weights
├── tokens/
│   └── obj_remesh/                 # Phase 2 output
│       ├── <id>_H.pt               [1024, 1024]  shape tokens
│       └── <id>_T.pt               [K, 1024]     skeleton tokens
├── docs/
│   ├── phase_1.md
│   └── phase_2.md
├── dataset.py                      # Phase 1: mesh → point cloud + skeleton NPY
├── tokenizer.py                    # Phase 2: point cloud + skeleton → tokens
├── tests/
│   ├── phase1_test.py              # Phase 1 validation + visualisation
│   └── phase2_test.py              # Phase 2 validation
├── CLAUDE.md
└── README.md

---

Quick Start

Environment

python -m venv .venv
source .venv/bin/activate
pip install open3d numpy

Phase 1 — Generate Point Clouds

# Quick test (5 shapes)
python dataset.py --max_shapes 5

# Full run (all 2703 shapes, resumes safely if interrupted)
python dataset.py --split all --max_shapes 2703 --resume

Phase 1 — Validate & Visualise

# Runs all checks + saves PNG renders to tests/output/
python tests/phase1_test.py

# Test a specific shape
python tests/phase1_test.py --shape_id 10000

See docs/phase_1.md for a full walkthrough.

---

Phase 2 — Tokenize Shapes & Skeletons

# Full run (all 2703 shapes, resumes safely if interrupted)
python tokenizer.py --max_shapes 2703 --resume

Phase 2 — Validate

python tests/phase2_test.py

See docs/phase_2.md for a full walkthrough.

---

Requirements

• Python 3.10+
• open3d, numpy