3D Arena: image-to-3D outputs of 29 generators on 100 shared input images

pl.load("hf_3d_arena") loads this dataset. The benchmark behind the 3D Arena Space: for each of ~100 input images (inputs/images, from the iso3D dataset) the output of 29 image-to-3D models under outputs//.glb|.ply with a rendered preview .png. Meshes are glTF (Y-up, metres by spec, scale is the generator's own); the .ply outputs of the Gaussian-splat generators (SAM-3D-Objects-3DGS) are splats. A sample is one (model, image) pair.

import pardesline as pl

ds = pl.load("hf_3d_arena")
print(ds.describe())  # roles, frames, units, labels, licence, citation
s = ds[0]
obj = s.mesh  # fetched now, cached under ~/.cache/pardesline/datasets/hf_3d_arena

Provenance and licence

Source	Hugging Face Hub repo 3d-arena/3d-arena, paths outputs
Scheme	hf
Size of the whole source	19.5 GB
Licence	MIT
Redistribution	unknown
Licence note	MIT on the Hub card (checked 2026-10-13). Outputs are generated by third-party models from iso3D input images; the input images' and each generator's terms are not stated here, TO CONFIRM before any use beyond evaluation.
Terms	https://huggingface.co/datasets/3d-arena/3d-arena
Homepage	https://huggingface.co/spaces/dylanebert/3d-arena
Cite	Ebert D. 3D Arena: An Open Platform for Generative 3D Evaluation. arXiv:2506.18787, 2025.
Samples grouped by	name

Roles (what each sample contains)

Role	Type	Frame	Units	Labels / layout	Pattern or file
mesh	mesh	Y-up	m		outputs/{model}/{name}.glb Generated glTF mesh (Y-up, metres by the glTF specification; absolute scale is arbitrary)
splat	splat	from the file header	from the file header		outputs/{model}/{name}.ply Generated PLY: Gaussian splats for the 3DGS generators (SAM-3D-Objects-3DGS); check the model folder for others
preview	image	from the file header	from the file header		outputs/{model}/{name}.png Rendered preview of the output

Standardized layout

Files are fetched on first access into the dataset cache and stay in their original format; ds.export(folder) writes every sample as standard files (PLY for points and meshes, MetaImage for volumes and label maps) with a .pardesline.json sidecar carrying frame, units, labels, licence and provenance, plus a manifest for the folder. Label maps become exact surfaces with to_mesh(label=k), DICOM series become one volume, raw arrays are decoded through the layout above.

ds = pl.load("hf_3d_arena").take(2)
ds.export("./standardized")        # PLY / MetaImage + sidecars + manifest.json

Algorithms and models that run on it

• image roles: dinov3_features.py (DINOv2 / DINOv3 patch features)
• mesh roles: mesh.to_points(num_points), mesh.decimate(), mesh.smooth(), mesh.clean(), mesh.to_open3d() ; catalog modules: point_cloud_processing, mesh_processing, voxel_to_mesh, sdf, point_cloud_reconstruction, registration, deformable_registration, point_cloud_segmentation, stereo_3d_reconstruction, lidar_scene_detection, lidar_camera_fusion, features_extraction, collision_detection, pyvista_mesh_processing, pyvista_pointcloud_processing, pyvista_volume_processing, pyvista_analysis
• splat roles: splat.to_numpy() (means), the gaussian-splatting guide ; catalog modules: point_cloud_processing, mesh_processing, voxel_to_mesh, sdf, point_cloud_reconstruction, registration, deformable_registration, point_cloud_segmentation, stereo_3d_reconstruction, lidar_scene_detection, lidar_camera_fusion, features_extraction, collision_detection, pyvista_mesh_processing, pyvista_pointcloud_processing, pyvista_volume_processing, pyvista_analysis
• Dense features: DINOv2 / DINOv3 on the image roles.

Related

• Datasets catalog · How to download a medical imaging dataset in Python · Add a dataset
• Tags: mesh, splat, image-to-3d, benchmark, hf

_{Generated from src/pardesline/data/manifests/hf_3d_arena.json on 2026-10-14.}