GenSAGA-3D / README.md
TrNi's picture
Update README.md
9e3dd84 verified
|
Raw
History Blame Contribute Delete
19.9 kB
metadata
language:
  - en
license: apache-2.0
task_categories:
  - text-to-3d
  - image-feature-extraction
  - zero-shot-classification
tags:
  - 3d-generation
  - text-to-3d
  - game-assets
  - safety
  - risk
  - safety-labels
  - content-safety
  - genre-taxonomy
  - genre-conditioned
  - autoregressive
  - mesh
  - point-cloud
  - 3d-gaussian-splatting
  - multi-view
  - dino
  - clip
size_categories:
  - 1K<n<10K
pretty_name: GenSAGA-3D

GenSAGA-3D

Genre-Conditioned, Safety-Aware 3D Game Assets for Text-to-3D Research

GenSAGA-3D is an open research dataset designed to study a part of text-to-3D generation that is still poorly represented in existing benchmarks: autoregressive 3D generation in structured game-asset space, with contextual safety labels and reusable multi-modal 3D representations.

The dataset contains 9,307 text-conditioned triangle meshes generated from 9,419 curated prompts, organized over 14 game genres × 29 functional asset-type groups, with 4 contextual safety tiers ranging from unambiguously safe to adversarial. Each asset is sanity-tested and distributed with a consistent set of water-tight geometry, multi-view, point-based, signed-distance, and embedding representations.

Why GenSAGA-3D?

Most existing text-to-3D evaluation datasets concentrate on diffusion-based generators and primarily organize samples around generic object categories or benchmark prompts. GenSAGA-3D is designed around four dimensions that are useful for studying the next generation of 3D foundation models:

  1. Autoregressive generation — a public corpus built from an autoregressive text-to-3D generator rather than a per-scene optimization or diffusion-only pipeline.
  2. Genre-conditioned structure — prompts are systematically organized by game genre and functional asset type, enabling controlled analysis across visual and gameplay contexts.
  3. Contextual safety — four safety tiers explicitly distinguish ordinary content, content that is acceptable within a game genre, borderline/dual-use content, and adversarial content.
  4. Research-ready 3D representations — every sample is accompanied by watertight geometry, SDF samples, a 3DGS-ready point cloud, calibrated multi-view renders, and pre-computed DINOv2/CLIP features.

Together, these properties make GenSAGA-3D useful not only as a generation benchmark, but also as a shared resource for 3D representation learning, multi-view understanding, safety modeling, retrieval, classification, and downstream 3D foundation-model research.


At a Glance

Property GenSAGA-3D
3D Assets 9,307 x 5
Game genres 14
Functional asset groups 29
Genre × asset-type prompt cells 113
Complexity tiers 3 (Basic / Refined / Complex)
Safety tiers 4 (T0–T3)
Watertight meshes 100%
2-manifold meshes 100%
Multi-view renders 18 / asset
Render resolution 800 × 800
SDF samples 200K / asset
3DGS-ready points 100K / asset
DINOv2 features 18 × 1024 / asset
CLIP features 18 × 768 / asset
Total storage ~178 GB
License Apache 2.0

Core Contributions

1. A public autoregressive text-to-3D corpus

GenSAGA-3D provides a large, openly released collection of text-conditioned meshes produced by an autoregressive 3D generator. This complements the predominantly diffusion-based landscape of public T2-3D evaluation resources and creates a concrete testbed for studying how autoregressive models behave across diverse prompts, object classes, genres, and safety contexts.

2. A structured genre × asset taxonomy

Instead of treating game objects as an unstructured collection of meshes, GenSAGA-3D defines a controlled taxonomy spanning 14 game genres and 29 functional asset groups, yielding 113 genre/asset-type prompt cells.

This structure supports controlled studies such as:

  • generation quality across genres;
  • geometry/topology changes as a function of genre;
  • performance across functional object categories;
  • cross-genre retrieval and classification;
  • systematic evaluation of semantic and stylistic coverage.

3. Contextual safety labels for 3D content

GenSAGA-3D uses a four-tier safety scheme:

Tier Meaning Typical interpretation
T0 Unambiguously safe Ordinary objects and benign game assets
T1 Safe-in-genre Content appropriate in an explicitly fictional/game context
T2 Borderline / dual-use Content whose risk depends on realism or context
T3 Adversarial Explicitly harmful, provocative, or adversarial content

The key design goal is contextual safety rather than binary safe/unsafe labeling. For example, a fantasy longsword may be appropriate as an RPG asset while a realistic weapon prompt in another context may require a different risk interpretation.

Safety tiers are assigned from the genre × asset-type construction rules; human validation is planned for a future release.

4. A multi-modal 3D research resource, not just a mesh dump

Every asset is packaged through a consistent four-layer representation stack:

  • Geometry: watertight triangle mesh.
  • Implicit / point representations: 100K surface samples + 100K volume samples with signed-distance values, plus a 100K-point 3DGS-ready cloud.
  • Multi-view supervision: 18 calibrated 800×800 views containing RGBA, normals, depth, and camera-to-world matrices.
  • Semantic embeddings: per-view DINOv2-Large and CLIP-ViT-L/14 features.

This allows researchers to use the same corpus for geometry, rendering, multi-view learning, semantic retrieval, classification, 3D Gaussian Splatting initialization, safety modeling, and multimodal representation research without rebuilding the preprocessing pipeline.

5. Strong geometric validity at release time

All 9,307 meshes are watertight and 2-manifold directly from the generator, with no repair needed during preprocessing. This provides a clean geometry substrate for downstream research and removes a common confound between generation quality and mesh-repair heuristics.


What Makes the Dataset Distinct?

GenSAGA-3D is intended to occupy the intersection:

                 AUTOREGRESSIVE 3D GENERATION
                            │
                            │
               ┌────────────┼────────────┐
               │            │            │
               ▼            ▼            ▼
        GENRE STRUCTURE  SAFETY      MULTI-MODAL 3D
               │        CONTEXT        REPRESENTATIONS
               └────────────┼────────────┘
                            ▼
                      GenSAGA-3D

The resulting dataset is useful for questions such as:

  • Does an autoregressive T2-3D generator generalize uniformly across game genres?
  • How does geometric complexity vary with genre and functional asset type?
  • Can 3D safety classifiers distinguish contextual game content from genuinely adversarial content?
  • Can multi-view visual features predict semantic alignment or safety without requiring full mesh processing?
  • How well do geometry, rendering, and embedding representations agree for the same generated asset?

Key Statistics

Metric Value
Total assets 9,307
Unique prompts 9,419 (112 duplicates across cells)
Generator Cube3D v0.5 (autoregressive; Apache 2.0)
Genres 14 (10 aesthetic + 4 gameplay-context)
Asset-type groups 29
Prompt cells (genre × asset type) 113
Complexity tiers Basic (968) / Refined (4,339) / Complex (4,000)
Safety tiers T0 (5,302) / T1 (2,617) / T2 (277) / T3 (1,111)
Watertight from generator 100%
2-manifold from generator 100%
Mean vertices 148,623
Mean faces 297,246
Multi-view renders 18 per asset, 800×800, 40° FOV
DINOv2 embedding 18 × 1024 float32 per asset
CLIP embedding 18 × 768 float32 per asset
Total storage ~178 GB
License Apache 2.0

Dataset Structure

GenSAGA-3D/
├── README.md
├── metadata.parquet          # One row per asset; complete metadata
├── combined_all.csv          # Human-readable CSV
├── combined_G{1..14}.csv     # Per-group CSVs
└── data/
    ├── 3dgs/                 # {asset_name}_3dgs.ply
    ├── wt/                   # {asset_name}_wt.npz
    ├── pts/                  # {asset_name}_pts.npz
    ├── views/                # {asset_name}_views.npz
    └── embed/                # {asset_name}_embed.npz

Selective download

The dataset is designed for modality-specific research. File sizes range from ~118 KB to tens of MB per asset. Use metadata.parquet to filter the corpus and download only the artifact families required for your experiment.

from huggingface_hub import snapshot_download

REPO = "YOUR_USERNAME/GenSAGA-3D"

# Metadata + pre-computed embeddings
snapshot_download(
    repo_id=REPO,
    repo_type="dataset",
    allow_patterns=["metadata.parquet", "combined_all.csv", "data/embed/*"],
)

# Watertight meshes only
snapshot_download(
    repo_id=REPO,
    repo_type="dataset",
    allow_patterns=["metadata.parquet", "data/wt/*"],
)

Metadata Schema

metadata.parquet / combined_all.csv contains 83 columns per asset.

Prompt and taxonomy

Column Type Description
prompt_hash str SHA-256 hash of the prompt text
asset_name str Unique asset identifier, e.g. G1A1_0
csv_row int Original prompt-table row
cell_id str Genre × asset-type prompt cell
genre str One of 14 game genres
asset_group str One of 29 functional asset groups
sub_type str Specific object subtype
material str Material descriptor
condition str Condition descriptor
size_tier str small / standard-sized / large / colossal
complexity_tier str Basic / Refined / Complex
safety_tier str T0 / T1 / T2 / T3
allocation_tier str Generation-planning allocation category
novel_flag bool Whether the prompt was marked as novel
prompt_text str Full text prompt supplied to the generator

Generation and geometry

Column Type Description
n_verts_gen int Vertex count from the generator
n_faces_gen int Face count from the generator
latency_s float Generation latency in seconds
status_render str Generation/rendering status

The repair/topology fields record both pre- and post-processing measurements. For this release, tier_used is none for all assets because no mesh repair was required.

Multi-view and embeddings

Column Type Description
n_views int 18
image_size int 800
camera_distance float 2.2
camera_fov_deg float 40.0
status_embed str Embedding generation status
path_3dgs_ply str 3DGS point-cloud path
path_wt_npz str Watertight mesh path
path_pts_npz str SDF/sample path
path_views_npz str Multi-view render path
path_embed_npz str Embedding path

Artifact Formats

_3dgs.ply — 3D Gaussian Splatting-ready point cloud

Binary little-endian PLY with 100,000 points and nine per-point properties:

Property Type Description
x, y, z float32 Normalized position
nx, ny, nz float32 Barycentric-interpolated normal
red, green, blue uint8 Placeholder color (128, 128, 128)

_wt.npz — watertight triangle mesh

vertices            (V, 3) float32
faces               (F, 3) int32
shift               (3,)   float32
scale               ()     float32
tier_used           ()     string
resolution_used     ()     int64

Meshes are normalized to a common coordinate convention.

_pts.npz — surface and volume samples

surface_points      (100K, 3) float32
surface_normals     (100K, 3) float32
surface_face_index  (100K,)   int32
surface_barycentric (100K, 3) float32
volume_points       (100K, 3) float32
volume_sdf          (100K,)   float32

_views.npz — calibrated multi-view supervision

rgba                (18, 800, 800, 4) uint8
normal              (18, 800, 800, 3) uint8
depth               (18, 800, 800) float32
camera_to_world     (18, 4, 4) float32
camera_angle_x      () float32

The 18 views are rendered from a Fibonacci-sphere camera configuration.

_embed.npz — frozen semantic features

dino_embeddings     (18, 1024) float32   # DINOv2-Large
clip_embeddings     (18, 768)  float32   # CLIP-ViT-L/14

Safety Tier Definitions

Tier Name Description Count
T0 Unambiguously Safe Benign objects with no intended harmful interpretation 5,302 (57.0%)
T1 Safe-in-Genre Potentially harmful-looking objects whose meaning is acceptable within the explicit game/genre context 2,617 (28.1%)
T2 Borderline / Dual-Use Objects whose interpretation depends on realism, intent, or context 277 (3.0%)
T3 Adversarial Explicitly harmful, provocative, or adversarial content 1,111 (11.9%)

Safety tiers are derived from the dataset's genre × asset-type cross-mapping rules. The rule basis is documented in the dataset construction methodology. Human validation is planned for v2.


Genre Taxonomy

10 aesthetic genres

ID Genre Example content
G1 Fantasy / Medieval Swords, shields, spell tomes, treasure
G2 Sci-Fi / Space Plasma weapons, spaceships, control panels
G3 Horror / Gothic Broken weapons, skeletons, coffins, medical instruments
G4 Steampunk / Dieselpunk Brass firearms, clock towers, steam engines
G5 Cyberpunk / Neo-noir Chrome katanas, neon storefronts, cybernetic implants
G6 Eldritch / Cosmic Horror Ritual totems, necronomicons, cosmic artifacts
G7 Post-Apocalyptic Pipe rifles, irradiated beasts, makeshift carts
G8 Mecha / Robotics Mech swords, armor plates, hover tanks
G9 Anime / Stylized Spirit staffs, mascot creatures, cosplay outfits
G10 Modern / Realistic Shotguns, headphones, IV stands, tennis rackets

4 gameplay-context genres

ID Genre Example content
G11 RPG / Adventure Treasure keys, scrolls, remedy vials, relics
G12 Combat / FPS Assault rifles, plate carriers, balaclavas
G13 Survival Ration packs, snare traps, bandages, makeshift carts
G14 Roleplay / Social Bookshelves, pizza, fruit bowls, guitars, hats

Example Uses

Safety-conditioned analysis

import pandas as pd

df = pd.read_parquet("metadata.parquet")

# Adversarial subset
T3 = df[df["safety_tier"] == "T3"]

# Safe-in-genre subset
T1 = df[df["safety_tier"] == "T1"]

print(len(T3), len(T1))

Genre-conditioned analysis

fantasy_melee = df[
    (df["genre"] == "Fantasy / Medieval") &
    (df["asset_group"] == "Weapons - Melee")
]

print(f"Fantasy melee assets: {len(fantasy_melee)}")

Load a watertight mesh

from huggingface_hub import hf_hub_download
import numpy as np

path = hf_hub_download(
    repo_id="YOUR_USERNAME/GenSAGA-3D",
    filename="data/wt/G1A1_0_wt.npz",
    repo_type="dataset",
)

mesh = np.load(path)
print(mesh["vertices"].shape, mesh["faces"].shape)

Research Impact

GenSAGA-3D is intended as a shared experimental substrate rather than a single-purpose benchmark. Its combination of structured prompts, safety labels, validated geometry, calibrated multi-view observations, and frozen visual embeddings enables several research directions:

  • Autoregressive T2-3D evaluation: compare generation behavior across genres, asset types, and complexity levels.
  • 3D safety: train and evaluate geometry-, image-, and multimodal safety classifiers.
  • Context-aware safety: study the difference between ordinary, safe-in-genre, borderline, and adversarial content.
  • Multi-view representation learning: use consistent 18-view observations with camera calibration.
  • 3D retrieval and classification: exploit meshes, renderings, DINOv2 features, CLIP features, or combinations.
  • Geometry analysis: investigate topology, complexity, manifoldness, and genre-dependent geometric statistics.
  • 3DGS and implicit representations: directly consume the released point clouds and SDF samples.
  • Efficient downstream experimentation: use the pre-computed representations without regenerating meshes or running GPU-heavy embedding pipelines.

Why this matters

A useful 3D benchmark should make it possible to separate generation quality, semantic coverage, contextual safety, and representation quality rather than conflating them in a single score. GenSAGA-3D is designed around that separation.


Positioning Relative to Existing 3D Resources

GenSAGA-3D is complementary to established 3D resources such as ShapeNet and Objaverse, which are primarily general-purpose repositories, and to T2-3D evaluation benchmarks that focus on generated-quality assessment.

Its distinguishing design choices are:

Dimension GenSAGA-3D
Generation paradigm Autoregressive text-to-3D
Prompt organization Genre × functional asset type
Safety Four-tier contextual labels
Geometry Watertight, 2-manifold meshes
Multi-view supervision 18 calibrated views
Implicit samples 200K surface/volume samples
3DGS representation 100K-point PLY per asset
Frozen visual features DINOv2 + CLIP
Licensing Apache 2.0

The goal is not to replace general-purpose 3D repositories, but to provide a structured experimental setting for generation, safety, and multimodal 3D analysis.


Limitations and Responsible Use

  • The assets are AI-generated, not scans of real-world objects.
  • The dataset reflects the characteristics and biases of the chosen autoregressive generator and the prompt taxonomy.
  • Safety labels are rule-based in the current release; human validation is planned for v2.
  • T1 explicitly represents contextual acceptability and should not be treated as universally safe outside the game-asset context.
  • Some T3 examples contain potentially sensitive or adversarial content and should be handled according to the policies of the downstream application.

Citation

@misc{gensaga3d2026,
  title        = {GenSAGA-3D: Genre-Conditioned, Safety-Aware 3D Game Assets for Text-to-3D Research},
  author       = {TBD},
  year         = {2026},
  howpublished = {Hugging Face Dataset},
  url          = {https://huggingface.co/datasets/YOUR_USERNAME/GenSAGA-3D}
}

License

GenSAGA-3D is released under the Apache License 2.0. See LICENSE for details.

A part of the dataset has been generated using the open, autoregressive, text-to-3D model Cube3D v0.5, which is also released under Apache 2.0. Cube3D is referenced solely for generator provenance and reproducibility; GenSAGA-3D is an independent dataset release and is not affiliated with or endorsed by Roblox.


Acknowledgments

We acknowledge the creators of the open components used in the dataset pipeline, including Cube3D v0.5, DINOv2-Large, and CLIP-ViT-L/14. These components are cited for reproducibility and technical provenance.