This corpus is part of BIMScript (paper · project page), where it supervises the per-element material and condition tokens. It was generated with the open-source pipeline at github.com/BIMScriptWorld/ASE-Material-Passport.

For every make_wall, make_door and make_window command in a scene's ase_scene_language.txt, we add:

material_type, from a closed vocabulary per class (see Vocabulary)
condition: new, good, worn or damaged

How the labels were made

Each element's surface is sampled in 3D and projected into the scene's posed RGB frames. Each projection is checked for visibility against the depth and instance renders, and the visible pixels are cut into texture crops. Three material classifiers then score the crops:

vlm: a vision-language model, google/gemma-4-E4B-it
dms: a SigLIP MINC material classifier, prithivMLmods/Minc-Materials-23
cv: a colour/texture heuristic

Their outputs are combined by priority consensus, in the order vlm → dms → cv. Condition comes from the same VLM call, with a luminance/texture rule as fallback.

In practice the VLM decides 99.6% of material and condition labels. The dms and cv predictions are kept in material_passport.json as *_candidates, so you can measure agreement or re-run the consensus.

Files

File	Size	Contents per scene
`ASE_with_Material_Passport_Slim.tar.gz`	545 MB	`ase_scene_material_lang.txt`, `material_passport.json`
`ASE_with_Material_Passport_Full.tar.gz`	TODO	Slim + `material_cache/`: the PNG texture crops the classifiers saw

Both archives extract to material_ase_lang_output/<scene_id>/, where <scene_id> is the original ASE scene id.

pip install -U huggingface_hub
hf auth login   # required: the dataset is gated
hf download prakashknaikade/ASE-Material-Passport \
  ASE_with_Material_Passport_Slim.tar.gz --repo-type dataset --local-dir .
tar -xzf ASE_with_Material_Passport_Slim.tar.gz

To get the ASE geometry, images and trajectories for the same scene ids, use the downloader in the pipeline repository (you need to accept the ASE license on the ASE website).

Missing scene ids. 20 of the 100,000 training scenes are not included:

9920–9929: the ASE chunk failed to download
9935, 9936, 9938, 25565, 26197, 26835, 27489, 28129, 28781, 29429: processing did not finish

`ase_scene_material_lang.txt`

The original ASE scene language, with two parameters appended to each structural line:

make_wall, id=0, a_x=-2.56, a_y=6.16, ..., thickness=0.0, material_type=painted_plaster, condition=good
make_door, id=1000, wall0_id=2, ..., height=2.02, material_type=composite (wood/metal), condition=good
make_window, id=2000, wall0_id=0, ..., height=2.12, material_type=aluminum, condition=worn

For composite materials, the constituents follow in parentheses (see Limitations).

`material_passport.json`

One record per structural element. line_index points to the line in ase_scene_material_lang.txt.

{
  "scene_dir": "0",
  "entities": [
    {
      "id": "wall0", "class": "wall", "command": "make_wall", "line_index": 0,
      "seg_id": 13, "n_samples": 40165, "observed": true,
      "material": "painted_plaster", "material_method": "vlm",
      "material_confidence": 0.95, "material_components": [],
      "material_candidates": [{"method": "dms", "label": "wood", "confidence": 1.0}, "..."],
      "condition": "good", "condition_method": "vlm", "condition_confidence": 0.95,
      "condition_candidates": [{"method": "vlm", "label": "good", "confidence": 0.95}, "..."],
      "features": {"mean_rgb": ["..."], "luminance_mean": 85.47, "texture_laplacian": "...", "...": "..."},
      "crops": ["material_cache/wall0_f0000038.png", "..."],
      "frames": []
    }
  ]
}

Field	Meaning
`observed`	`false` if no frame showed the element (0.36% of elements). Its `material` and `condition` are then `null`, and `*_method` is `"none"`.
`material_method` / `condition_method`	Backend that won the consensus.
`material_components`	Constituents when `material` is `composite`, otherwise empty.
`*_candidates`	Every backend's prediction, with its confidence.
`features`	Colour and texture statistics of the sampled pixels.
`crops`	Paths of the crops in `material_cache/`. The files are in the Full archive only.

Vocabulary

wall: painted_plaster, wallpaper, brick, concrete, wood_paneling, tile, stone, wood, metal, glass, composite
door: wood, metal, glass, composite
window (frame material; glazing is always glass): wood, aluminum, pvc
condition: new, good, worn, damaged

Statistics

Class	Elements	Most common materials	Condition (good / new / worn / damaged)
wall	1,133,283	wallpaper 31.9%, painted_plaster 23.4%, wood_paneling 22.1%, tile 10.6%, concrete 4.2%	80.5 / 16.3 / 2.9 / 0.1 %
door	260,755	composite 65.2%, wood 24.1%, metal 10.2%, glass 0.2%	80.8 / 17.7 / 1.2 / 0.0 %
window	506,870	aluminum 77.0%, wood 21.4%, pvc 1.0%	89.7 / 2.8 / 6.9 / 0.0 %

Percentages are of all elements in the class. The remainder are unobserved elements: 0.2% of walls, 0.3% of doors and 0.6% of windows.

Limitations

Pseudo-labels. The labels come from a single small VLM and have not been verified by humans, so expect noise. Condition is especially uncertain: on clean synthetic renders it is a plausibility estimate rather than real wear, and worn/damaged are rare.
Unobserved elements. In the JSON, unobserved elements have null labels. In ase_scene_material_lang.txt they get a fixed class default instead (wall painted_plaster, door wood, window aluminum, condition good). Use observed in the JSON to filter them out.
Composite constituents are free text. The parts listed in parentheses come straight from the VLM and are not normalised. They can include terms outside the vocabulary (plastic, resin, fiberboard), appear in either order (wood/metal, metal/wood), or name composite itself (about 21% of composite labels, e.g. metal/composite). Treat composite as the label, and the parts as a hint.
Synthetic domain. Labels describe the rendered ASE textures and may not transfer to real-world imagery.

License and terms of use

This dataset is derived from the Aria Synthetic Environments dataset and is provided for non-commercial research purposes only. All use of it, including downstream use and redistribution, is governed by the Aria Synthetic Environments Dataset License Agreement. Publications using this data must include the attribution: [Aria Synthetic Environments Dataset, Meta Reality Labs-R].

The pipeline code is MIT-licensed; see the GitHub repository.

Citation

If you use this dataset, please cite BIMScript:

@misc{naikade2026bimscript,
  title         = {BIMScript: Material-Aware Structured Scene Programs for BIM Ingestion},
  author        = {Naikade, Prakash Kondibhau and Moeslund, Thomas B. and M{\o}gelmose, Andreas},
  year          = {2026},
  eprint        = {2608.21447},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2608.21447}
}

Please also credit the underlying ASE data with the attribution required by its license: [Aria Synthetic Environments Dataset, Meta Reality Labs-R].

Downloads last month: 72

Total file size:

545 MB

Paper for prakashknaikade/ASE-Material-Passport

BIMScript: Material-Aware Structured Scene Programs for BIM Ingestion

Paper • 2608.21447 • Published Aug 19