You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Splataverse-T

A collection of synthetic environments represented as tiled 3D Gaussian splats, with available material channels, labels and per-Gaussian face/barycentric correspondences to the source meshes. Scenes cover indoor spaces, industrial environments, streets and natural landscapes.

Public, manually gated release. Request access on this page and sign in with an approved Hugging Face account to download the files.

Current status: partial release — 59 remotely verified scenes from a target inventory of 66. The manifest distinguishes verified uploads from scenes still being processed. A conversion finishing locally does not imply that its files are available here.

Scene preview

Scene View 1 View 2
OldTown OldTown view 1 OldTown view 2

Download a scene

Install huggingface_hub and tyro, then authenticate:

pip install -U huggingface_hub tyro
hf auth login
hf download maxwildersmith/splataverse-t download_scene.py archive_transport.py --repo-type dataset --local-dir splataverse-t-tools
python splataverse-t-tools/download_scene.py --scene OldTown --output-dir ./OldTown

The downloader pins the selected repository revision, supports both loose files and archive-backed scenes, and verifies original payload hashes. Use --revision COMMIT to request a particular release. Select a scene marked VERIFIED in manifest.json.

Contents and quality scope

Each available scene contains tiled Gaussian splats (tiles/*.spz) with per-Gaussian face/barycentric correspondences (*.correspondence.npz), material fields (*.pbr.npz) and semantic labels (*.semantic.npz), plus camera, density and conversion reports. Availability is recorded per scene; unevaluated or unavailable material channels are not claimed as ground truth.

New conversions use an analytical balanced CPU profile: one source-based sampling plan and one emission, with per-instance geometry scale and source-material detail. Existing exports retain their original versioned balanced recipe and validation records. The new analytical variant is published separately after qualification; existing scene bytes are not relabeled or overwritten.

Rendering and LPIPS evaluation are separate from conversion. Records marked CPU_ONLY_PREVIEWS_DEFERRED have no rendered appearance or LPIPS measurement. Earlier source-export measurements retain their actual scores and view scope; they do not establish general equivalence to arbitrary native engine shaders or held-out material inference. The native engine reference renders used for those measurements are not redistributed.

Storage and provenance

Scenes marked exact_payload_tar_shards_v1 use lossless TAR transport under archive_release/v1/scenes/. Archive extraction reconstructs the shipped logical files and verifies their hashes. Previously published loose files remain available.

The source scenes are not redistributed: no Unreal packages, source textures, USD meshes or source geometry index maps are included. Provenance is provided through the face/barycentric correspondences, which identify the source mesh face and position for every Gaussian, together with the source identities, asset paths, attribution and license notices recorded at export time. Each quality.json lists only the shipped files and records the removed source files under source_redistribution; the shipped files are byte-identical to the original conversion outputs, and the downloader verifies their hashes. Machine-specific paths are omitted from the public inventory. The repository history was squashed when the source files were removed, so earlier commit ids are no longer available. Source-scene license terms are recorded individually and remain unverified where terms were not supplied; no blanket redistribution license is asserted. Manual access approval does not replace applicable source terms.

This synthetic collection is separate from the real-city dataset index.

Downloads last month
436