README.md

---
pretty_name: ProSceneverse
license: other
size_categories:
  - 10K<n<100K
task_categories:
  - text-to-3d
  - image-to-3d
  - other
tags:
  - 3d
  - 3d-scenes
  - glb
  - gltf
  - sketchfab
  - captions
  - scene-understanding
language:
  - en
annotations_creators:
  - machine-generated
source_datasets:
  - original
configs:
  - config_name: default
    data_files:
      - split: train
        path: metadata.parquet
---

# ProSceneverse

Hand-designed 3D scenes (CAD, game environments, designed interiors and
exteriors, urban scenes, stylized landscapes) curated from the TexVerse-1K
Sketchfab crawl — the companion of
[ScanSceneverse](https://huggingface.co/datasets/Stable-X/ScanSceneverse) which
holds real-world photogrammetry scans.

**74,624** textured designed scenes in glTF/GLB format, each with a
machine-generated English caption, a scene-type label, and a thumbnail
render.

## Contents

```
ProSceneverse/
├── README.md
├── metadata.parquet          # 74,624 rows, one per scene
├── shard_index.parquet       # asset path -> tar shard it lives in
└── shards/
    ├── glbs/                 # glbs-0000.tar … glbs-0068.tar  (69 tars, ~15 GB each)
    └── thumbnails/           # thumbnails-0000.tar            (1 tar, ~1.8 GB)
```

Assets ship as **uncompressed tar shards**, not as 74,624 loose files — the Hub
rate-limits that many small objects to the point of stalling. Members inside
each tar keep the original layout (`glbs/000-NNN/<scene_id>.glb`,
`thumbnails/000-NNN/<scene_id>.jpeg`), so extracting all shards into one
directory reproduces exactly the tree that the `file_name` and `thumbnail`
columns of `metadata.parquet` point at.

## Categories

- `CAD_ARCHITECTURE`: **44,487**
- `GAME_SCENE`: **17,303**
- `DESIGNED_ROOM`: **7,183**
- `DESIGNED_EXTERIOR`: **2,539**
- `URBAN_SCENE`: **1,776**
- `DESIGNED_LANDSCAPE`: **1,336**

## Licenses

- `by`: 68,186
- `by-nc-sa`: 3,767
- `by-sa`: 1,083
- `by-nc`: 847
- `by-nc-nd`: 384
- `unknown`: 188
- `cc0`: 118
- `by-nd`: 37
- `free-st`: 14

## Metadata schema

| column | type | description |
| --- | --- | --- |
| `scene_id` | string | 32-hex Sketchfab model id (primary key) |
| `file_name` | string | Relative path to the GLB |
| `thumbnail` | string \| null | Relative path to the JPEG preview |
| `caption` | string | Gemini-generated English caption of the scene |
| `category` | string | Curated scene type |
| `decision` | string | Curation decision (`KEEP`) |
| `scanning_ids` | list<string> | `scanned_by_*` tags (empty for designed scenes) |
| `name` | string | Original Sketchfab model title |
| `description` | string | Original Sketchfab model description (author-written) |
| `tags` | list<string> | Author-provided Sketchfab tags |
| `sketchfab_categories` | list<string> | Original Sketchfab category names |
| `license_slug` / `license_label` / `license_url` | string | Per-scene license |
| `author_username` / `author_display_name` / `author_uid` | string | Original Sketchfab author |
| `face_count`, `vertex_count`, `material_count`, `texture_count`, `animation_count` | int64 | Geometry/material stats |
| `pbr_type` | string \| null | PBR workflow tag if set |
| `source` | string | Sketchfab upload source |
| `published_at`, `updated_at` | string (ISO-8601) | Timestamps from Sketchfab |
| `source_url` | string | Public Sketchfab viewer URL |
| `view_count`, `like_count`, `download_count`, `comment_count` | int64 | Popularity stats at crawl time |

## Loading

The metadata is a plain parquet file — no asset download needed to browse
captions, categories, or geometry stats:

```python
from datasets import load_dataset
ds = load_dataset("parquet", data_files="metadata.parquet", split="train")
print(ds[0]["caption"], ds[0]["file_name"])
```

### Getting the assets

Full dataset (~1.0 TB), extracted back into the original layout:

```bash
hf download Stable-X/ProSceneverse --repo-type dataset \
    --include "shards/**" --local-dir ProSceneverse
cd ProSceneverse && for t in shards/glbs/*.tar shards/thumbnails/*.tar; do tar -xf "$t"; done
# -> glbs/000-NNN/<scene_id>.glb, thumbnails/000-NNN/<scene_id>.jpeg
```

Only the scenes you care about — `shard_index.parquet` maps each asset to its
tar, so you can fetch a handful of shards instead of the whole set:

```python
import pandas as pd, tarfile
from huggingface_hub import hf_hub_download

idx = pd.read_parquet("hf://datasets/Stable-X/ProSceneverse/shard_index.parquet")
wanted = ["glbs/000-000/0a1b....glb"]                      # paths from metadata.parquet
for shard in idx[idx.path.isin(wanted)]["shard"].unique():
    local = hf_hub_download("Stable-X/ProSceneverse", shard, repo_type="dataset")
    with tarfile.open(local) as tf:
        tf.extractall(".", members=[tf.getmember(p) for p in wanted
                                    if p in tf.getnames()])
```

Then resolve meshes relative to the extraction root, e.g. with `trimesh`:

```python
import trimesh, os
root = "/path/to/ProSceneverse"
mesh = trimesh.load(os.path.join(root, ds[0]["file_name"]))
```

## Licenses

Every source model on Sketchfab ships with its own Creative Commons
license. The per-scene license is recorded in the `license_slug`,
`license_label`, and `license_url` columns — consult them before
redistribution.