Dataset Viewer
Auto-converted to Parquet Duplicate
The dataset viewer is not available for this split.
Server error while post-processing the rows. This occured on row 5. Please report the issue.
Error code:   RowsPostProcessingError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

NASA 3D Resources

A Hugging Face packaging of NASA 3D Resources (GitHub mirror): spacecraft, rovers, launch vehicles, spacesuits, tools, ground facilities, astrophysical objects, 3D-printable terrain and planetary texture maps produced across NASA centers.

This is not a complete mirror: some NASA assets are deliberately excluded.

  • NASA insignia. NASA does not permit its insignia (the "meatball", the "worm" logotype and the seal) "to appear with AI generated imagery or in the training of AI tools". Every asset in which screening found one is left out: 23 GLB files covering 14 models (Astronaut, Astronaut Glove, Gateway, Global Hawk, Habitat Demonstration Unit, InSight Cruise Lander, Ingenuity Mars Helicopter, ISS (D) (IGOAL), Mark III Spacesuit, Mars 2020 Perseverance Rover, Pistol Grip Tool, Space Shuttle (D), Vehicle Assembly Building, Z2 Spacesuit), the NASA Insignia and CubeSat printables (all of their parts), and 41 original source files across 12 subjects. See Insignia screening for how assets were screened.
  • Unavailable files. 7 files listed on science.nasa.gov return HTTP 404. Four of them are included from the GitHub repository instead; the three Dragonfly slicer project files (Bambu.3mf, Orca.3mf, Prusa.3mf) are not available anywhere and are missing.
  • Duplicates. 102 byte-identical copies (same SHA-256) of files that are already included are collapsed into a single row; the other titles they appeared under are kept in also_listed_under.

Every excluded file is listed with its reason in meta/excluded.jsonl.

Apart from the exclusions above, this mirror changes only the packaging. Files are grouped into typed Parquet shards and every original file is preserved byte-for-byte: GLB and STL files use the native datasets Mesh() feature, previews and texture maps use Image(), and every other original file (Blender, 3DS, LightWave, FBX, zip and 7z archives, print guides) is kept verbatim in the source_files config.

Configs

Config Row unit Rows Main fields
models (default) One GLB file 235 mesh (Mesh), preview (Image), title, description, collection, contributors, mesh stats
printables One STL part 432 mesh (Mesh), preview, part, archive (zip it came from, if any)
textures One image or texture map 99 image (Image, original JPG/TIF bytes)
source_files One other original file 165 blob (raw bytes), extension

models covers 206 distinct catalog subjects (0.39 GB of GLB). Some subjects ship several GLB variants; each variant is its own row.

from datasets import load_dataset

ds = load_dataset("suvadityamuk/nasa-3d-resources", split="train")          # models
row = ds[0]
row["title"], row["contributors"]
scene = row["mesh"]          # trimesh.Scene (pip install trimesh DracoPy)
row["preview"]               # PIL.Image

stl = load_dataset("suvadityamuk/nasa-3d-resources", "printables", split="train", streaming=True)

For the raw file bytes use ds.cast_column("mesh", Mesh(decode=False)), which yields {"bytes": <original file bytes>, "path": <file name>}.

Some GLBs use Draco mesh compression (draco_compressed = true). Install DracoPy alongside trimesh (pip install trimesh DracoPy); without it trimesh decodes Draco geometry as placeholder zeros.

Provenance and joins

Metadata (title, description, collection, contributor, publish dates, page URL) comes from the science.nasa.gov catalog (WordPress REST API). Files come from the catalog's download links and from the GitHub repository; when both offer the same file name the GitHub copy is used and source says so. Every row records source_url, github_path and sha256.

Data-quality notes inherited from the source:

  • Some catalog pages point at the wrong file or folder (for example the Rosetta page's download is the IBEX model, and the Voyager Probe (A) and Ares 1 (B) pages link the Viking Lander and Ares 1 (A) repo folders). Repo folders are matched by title first, and identical files are collapsed by sha256; the copy whose file name matches its title is kept, the others are listed in also_listed_under and in meta/excluded.jsonl.
  • file_name_matches_title is false where a file's name does not match its subject.
  • Nine catalog models have no GLB (only .blend, .3ds, .zip or split .7z files); their originals are in source_files and they are absent from models.

License and usage

NASA states that "NASA content – images, audio, video, and media files used in the rendition of 3-dimensional models, such as texture maps and polygon data in any format – generally are not subject to copyright in the United States" (NASA Images and Media Usage Guidelines). The repository README describes the assets as "free and without copyright"; its meta.json lists the NASA Open Source Agreement 1.3, a software license that does not fit these assets. This card therefore uses license: other and points to the guidelines. The copyright status outside the United States is not stated by NASA.

Please read the guidelines before use. In particular:

  • Credit. "NASA should be acknowledged as the source of the material."
  • No endorsement. NASA content may be used "in a factual manner that does not imply endorsement." This dataset contains NASA source material; NASA has not reviewed, approved or endorsed it.
  • Insignia. "The NASA Insignia, Logotype, identifiers, and imagery are not in the public domain," and for AI applications "NASA does not permit our Insignia (i.e., NASA Meatball, Worm logotype, or Seal) to appear with AI generated imagery or in the training of AI tools." This mirror therefore excludes every asset in which a NASA insignia was found (see Insignia screening).
  • AI outputs. The guidelines ask that AI outputs not be attributed to NASA ("according to NASA" is prohibited) and encourage marking AI-generated products as such.
  • Third-party contributors. Rows with third_party_contributor = true were contributed by organisations outside NASA (for example DigitalSpace Corporation, Made in Space, STScI, JHU APL, SwRI, Rutgers). They may not be U.S. Government works; check the contributors field and the source page.

Insignia screening

Every image a model could learn from was screened for the NASA meatball, worm logotype and seal with Qwen/Qwen3-VL-32B-Instruct, and every positive was reviewed by hand before deciding:

  1. every texture embedded in every GLB, every preview render, and every texture map (large images tiled so small logos are visible);
  2. every STL part rendered from six orthographic directions with relief shading, so embossed or engraved insignia are visible, and every image inside the zip / 7z archives;
  3. PNG/JPEG textures carved out of .blend, .fbx and other binary scene files, rendered PDF pages and PowerPoint media.

Excluded as a result (listed per file in meta/excluded.jsonl):

  • 23 GLB files of 14 subjects (for example Space Shuttle (D), ISS (D), InSight, Ingenuity, Perseverance, Z2 and Mark III spacesuits, Global Hawk, Vehicle Assembly Building);
  • the NASA Insignia and CubeSat printables (all parts of a printable object are dropped if any part carries an insignia);
  • original scene files and archives that contain insignia textures, and all original files belonging to the subjects above.

Screening is automated plus manual review and may miss very small or heavily stylised marks. Files that could not be screened (Maya .mb scenes with .iff textures, and .3ds / .lwo / .max / .vsp3 files without embedded raster images) remain in source_files; check them before using them to train generative models. The VLM verdicts are in the screening repo notes.

Attribution

Credit: NASA, via NASA 3D Resources (https://science.nasa.gov/3d-resources/). Individual contributor credits are in the contributors column.

Downloads last month
202

Space using hugging-3d/nasa-3d-resources 1