Astra3D-Bench-Trajectory

⚠️ Access requires permission. To use this dataset, please first email yansongning002@gmail.com to request permission, then submit the access request on this page. Access will be granted after your email has been reviewed.

⚠️ 使用需申请授权。 如需使用本数据集，请先发送邮件至 yansongning002@gmail.com 申请许可， 再在本页面提交访问申请，邮件审核通过后即开通访问权限。

Overview

Astra3D-Bench-Trajectory contains 1,000 3D scene-generation tasks (Q0001–Q1000) together with the agent trajectories and final deliverables produced while building each scene in Blender. Each task gives a natural-language scene description (Chinese), sometimes with reference images, and the agent constructs a complete 3D scene from it. For every task the dataset provides:

• the task query and a structured scene card used as evaluation metadata;
• the final scene as .blend and .glb, plus a self-contained browser viewer;
• rendered images and browser screenshots;
• validation records (browser loading check, Blender ↔ glTF round-trip check and edit);
• the full reasoning / interaction process (in the reasoning_split*.tar.gz archives).

Total size: about 106 GB.

Tasks

Two input modalities:

• text_to_3d: scene description only;
• image_to_3d: scene description plus reference image(s).

Each scene card annotates the task along several dimensions, including:

• application domain, e.g. museum / virtual exhibition, architecture & landscape design, film & animation production, game & interactive entertainment;
• scene space (indoor, outdoor urban, outdoor natural, …) and indoor subtype;
• layout, terrain, ecology, season, climate zone, geo-culture, historical period;
• visual style and complexity;
• functional zones, core elements, spatial relations, domain requirements, must_include, must_avoid.

All scenes are static. Motion described in the query is represented as a single frozen instant, with no animation or physics simulation.

Repository Structure

Q0001/
├── query.json               # task: query text, scene card, input modality, reference images, policies
├── final.blend              # final Blender scene
├── final.glb                # final scene exported to glTF binary
├── final_state.json         # final scene state
├── scene_report.json        # agent's summary: dimensions, implementation, assumptions, known differences
├── delivery.json            # delivery record with artifact paths and SHA-256 checksums
├── overview.png / detail.png / overall.png   # native Blender renders (file names vary by task)
├── viewer.html              # offline browser viewer for final.glb
├── OPEN_VIEWER.txt          # how to open the viewer locally
├── browser.png              # browser rendering screenshot
├── browser_interaction.png  # browser interaction screenshot
├── browser_json.png
├── browser_check.json       # browser loading / rendering check
├── browser_progress.json
├── roundtrip.blend          # .glb re-imported into Blender
├── roundtrip_edit.json      # edit applied in the round-trip test
├── roundtrip_edited.glb     # re-exported scene after the edit
├── roundtrip_check.json     # round-trip check result
└── PROVENANCE.txt           # source run and archive of this task
...
Q1000/
reasoning_split01.tar.gz … reasoning_split10.tar.gz   # agent reasoning processes (rollouts, commands, turns, renders)

Some task folders contain extra renders or intermediate files (for example per-area detail images).

Usage

After your access request has been approved, log in and download:

huggingface-cli login
huggingface-cli download yasNing/Astra3D-Bench-Trajectory --repo-type dataset --local-dir Astra3D-Bench-Trajectory

To download only a few tasks, without the large reasoning archives:

huggingface-cli download yasNing/Astra3D-Bench-Trajectory --repo-type dataset \
  --include "Q0001/*" "Q0002/*" --local-dir Astra3D-Bench-Trajectory

To view a scene in the browser:

cd Astra3D-Bench-Trajectory/Q0001
python3 -m http.server 8000 --bind 127.0.0.1
# then open http://127.0.0.1:8000/viewer.html

The viewer has no external dependencies (no CDN). .blend files can be opened in Blender, and .glb files in any glTF viewer.

License and Terms of Use

This dataset is released under the GPL-3.0 license. Access is subject to approval by the authors: please email yansongning002@gmail.com before requesting access. Do not redistribute the data without permission.

Contact

Yansong Ning: yansongning002@gmail.com