maxwildersmith/splataverse-simready

Splataverse SimReady

Uncapped, CPU-converted Gaussian representations of NVIDIA SimReady Assets, with source geometry/material correspondences and authored articulation metadata.

Release in progress. The first two published, independently validated assets contain 36,542,161 Gaussians. Additional assets are processed and published only after source-correspondence checks pass. This is not a claim that the entire source collection has finished conversion.

Initial validated asset	Gaussians	CPU conversion	Authored USD physics joints
case_d01	33,243,787	234.6 s	5
kitchenware_pan_n01	3,298,374	21.7 s	0

Times are the measured conversion stage on four Euler CPU cores, excluding input download, independent validation and upload. Neither conversion used a GPU. The case archive is 1.293 GB and the pan archive 151.9 MB. Every published archive is matched to a remotely verified SHA256.

Source and attribution

Original assets: NVIDIA, SimReady Assets, revision 7a0f37985996d244e49cedf92ff1df7e4424247e. The complete pinned inventory contains 2,632 USD-bearing root packages, 180,867 files and 481,543,637,192 bytes; the source card advertises 2,603 assets. A root-package count is not assumed to equal an independently audited object/assembly count.

The original per-package CC BY 4.0 license and source attribution are retained. Access approval is manual; this does not replace or narrow CC BY 4.0 rights.

Representation and provenance

Each Gaussian retains original triangle/polygon/corner/vertex IDs, barycentric coordinates, UVs, source mesh/material identity and a source rigid-body link binding where available. Source material channels include normal, roughness, metallic, emissive and opacity. Explicit UsdPreviewSurface inputs are sampled directly; unsupported or MDL-only materials are reported as failures rather than replaced with gray defaults.

The archive retains original USD/material/metadata files, pinned source URLs and hashes, exact sampling settings/tool versions/code hashes, authored typed properties, instance/prototype stacks, source transforms, units/up axis, variants, raw physics/PhysX/custom metadata, normalized joints and link-local Gaussian coordinates. Original textures remain available through pinned source references and checksums.

Coordinates are right-handed Y-up meters. SPZ packing uses explicit per-chunk units and origins: apply xyz_m = decoded_xyz / spz_units_per_meter + origin and scales_m = exp(decoded_log_scales) / spz_units_per_meter. Lossless source scales and quaternions are provided separately. See DATA_FORMAT.md.

Authored joints retain their original local frames, relationships, limits and drives. Missing authored joint states remain null; absence of joints is explicit. Validation tests two clearly labeled derived states within authored USD schema limits on an eligible joint, including link-bound positions, covariance eigenvalues and unchanged source IDs. Potential source unit inconsistencies are retained rather than silently reinterpreted.

Validation and inspection previews

Every released Gaussian passes original-source topology/barycentric reconstruction, finite/length checks, link binding, member/archive checksums and an independent SPZ-header-derived quantization bound. The first case has maximum source reconstruction error 1.25e-6 m and link reconstruction error 5.96e-8 m. These engineering checks do not measure appearance LPIPS.

The images below are CPU nearest-surface Gaussian-center inspection projections using all emitted points. They are not alpha-composited Gaussian renders, lit PBR results or evidence of measured image fidelity. PBR channels are included for downstream renderers.

case_d01	kitchenware_pan_n01

Files

assets/<asset_id>.tar contains Gaussian chunks, sidecars, source provenance, license and two inspection views. Matching .json and .validation.json files report actual counts, settings, hashes and independent validation. Conversion has no per-asset/global Gaussian count cap or image downsampling. Resource guards stop and record incomplete assets rather than reducing density or publishing partial results.