RedFairy/SceneRig-benchmark

SceneRig Benchmark

80 input images paired with SceneRig's settled Blender outputs and self-contained USDZ scenes, reconstructed with Claude Opus 5 (claude-opus-5).

The collection contains 44 simulator images, 18 real robot workspace images, and 18 smartphone, internet, or AI-generated images. These are reconstructed outputs, not ground-truth 3D annotations. All examples are in the test split.

Project page · Code

Files

images/<scene_id>.png              Original benchmark input
scenes/<scene_id>/settled.blend    Saved final Blender output, with packed textures
scenes/<scene_id>/scene.usdz       USD scene with physics, geometry, textures, lights, and camera
manifest.jsonl                    Image-to-scene mapping, provenance, and checksums
release_metadata.json             Packaging and validation information
SHA256SUMS                        File checksums
data/test-00000-of-00001.parquet   Dataset table with embedded input images

Each of the 80 examples has exactly one input image, one Blender file, and one USDZ file. image_path, settled_blend, and usdz are paths relative to this repository. The table also records the scene ID, source family, image dimensions, base model, rigid-body count, saved verification statuses, and asset checksums.

Source family	Images
RoboDojo simulator	24
RoboLab simulator	20
Robot workspaces	18
Other images	18
Total	80

Download and use

from datasets import load_dataset
from huggingface_hub import hf_hub_download

repo_id = "RedFairy/SceneRig-benchmark"
example = load_dataset(repo_id, split="test")[0]
image = example["image"]
blend_path = hf_hub_download(
    repo_id, example["settled_blend"], repo_type="dataset"
)
usdz_path = hf_hub_download(
    repo_id, example["usdz"], repo_type="dataset"
)

Download everything and verify its checksums:

hf download RedFairy/SceneRig-benchmark --repo-type dataset --local-dir SceneRig-benchmark
cd SceneRig-benchmark
sha256sum --check SHA256SUMS

Open settled.blend in Blender, or load scene.usdz in a USD-compatible application. Simulation requires a runtime that supports the authored USD Physics/PhysX schemas; a generic 3D viewer only displays the scene.

Provenance and validation

• This release uses the original 80 runs in static_scene_benchmark_final. Each run records claude-opus-5 as its model. Input image checksums match the preprocessing manifests.
• settled.blend is an unchanged copy of scene/final/final.blend, verified against the pipeline manifest. “Settled” refers to the pipeline's saved output after composition settling, not an additional simulation performed for this release. All 80 saved composition certifications report passed.
• USDZ files were generated from each run's saved scene/isaac/scene.usd using UsdUtils.CreateNewUsdzPackage, the packaging step documented in the project's Isaac export scripts. No reconstruction, simulation, or physics tuning was rerun.
• The 708 identity-mapped rigid-body transforms agree between the Blender and USD scenes within 1e-4 per matrix entry. USDZ packaging preserves those transforms exactly. Every package has resolved dependencies, uncompressed entries, and 64-byte file alignment.
• Blender files have packed textures and no linked-library dependencies; they were checked with Blender 4.5.13. Both native formats retain the saved Z-up scene convention. Materials and lighting may render differently between applications.
• The separate archived USD verifier reports dynamic_verified for 4 scenes and unstable for 76 under its recorded thresholds. These statuses are preserved in the table as usd_simulation_status; composition certification is a distinct check. This release does not claim that every scene passes that separate verifier.

The repository retains its Apache-2.0 license designation.