Cosmic-Gen Training Data
Curated 3D character + prop + environment data for training the Cosmic Corrector v3 model and downstream Cosmic-Gen pipeline modules.
Structure
cosmic-gen-data/
├── lists/
│ └── TexVerse_pbr_id_list.txt # 158,518 confirmed-PBR object IDs from TexVerse
├── scrapers/ # Reproducible scrapers for upstream sources
│ ├── sketchfab_scraper.py # CC0 models via Sketchfab API
│ ├── uitars_scraper.py # Agentic GUI scraper (bypasses anti-bot)
│ ├── cgtrader_playwright_scraper.py
│ ├── vroid_scraper.py
│ └── push_to_hf.py
├── docs/
│ ├── plan.md # Production plan
│ └── RESEARCH.md # Literature review + architecture decisions
├── sketchfab_cc0/ # (filled by Sketchfab scrape)
│ ├── metadata.jsonl
│ ├── glbs/<uid>.glb
│ └── thumbs/<uid>.jpg
└── cgtrader/ # (filled by UI-TARS scrape later)
Data Sources & Licenses
| Source | License | Count | Status |
|---|---|---|---|
| TexVerse PBR | ODC-By | 158,518 | IDs listed; download via YiboZhang2001/TexVerse |
| Sketchfab CC0 | CC0 | TBD (target ~10k) | Scraper ready |
| VRoid Hub | Varies (filter to permissive) | TBD | Scraper ready |
| CGTrader free | Various permissive | TBD | UI-TARS agentic scraper ready |
| Mixamo | Adobe free | 2,773 | via jasongzy/Mixamo HF mirror |
Usage
from huggingface_hub import snapshot_download
local = snapshot_download("PrimeStriker/cosmic-gen-data", repo_type="dataset")
# Then read lists/TexVerse_pbr_id_list.txt to filter TexVerse for PBR-only models
Cosmic_Gen project
This dataset is used to train modules of the Cosmic-Gen pipeline:
- Image → 3D character mesh + rigging
- Built on top of TRELLIS.2 (Microsoft), UniRig (CMU), SDXL (Stability)
- Our contributions: Cosmic Corrector v3, content classifier, agentic scraper
Maintainers
- @PrimeStriker (lead)
- ml-intern-explorers org
Citation
If you use this dataset, please cite the upstream sources (TexVerse, Sketchfab CC0, etc) as listed in metadata.jsonl of each subfolder.
- Downloads last month
- 586