leopoldmaillard/sceneteract-renders

SceneTeract Scene Renders

The images a VLM is shown when judging whether an activity is physically feasible in a 3D indoor scene. Pair these with sceneteract-traces to evaluate a model against geometry-grounded labels.

3,621 renders over 1,132 3D-FRONT living rooms and dining rooms, 1024×1024 PNG.

from datasets import load_dataset

renders = load_dataset("leopoldmaillard/sceneteract-renders")["train"]
renders[0]["image"]        # PIL image

Two kinds of image

kind	count	keyed by	used for
prompting_view	1,132	scene_id	task-level judgment — the whole activity, from a top-down view of the room
target_highlight	2,489	scene_id + object_id	action-level judgment — the same view with the target object hilighted by a red bounding box

column	type
scene_id	string	3D-FRONT scene, as preprocessed by PhyScene
room_type	string	LivingRoom or DiningRoom
kind	string	prompting_view or target_highlight
object_id	string	the highlighted object; null for prompting_view
image	Image	1024×1024 PNG (RGBA — convert to RGB before feeding a model)

Joining with the traces

from datasets import load_dataset

traces  = load_dataset("leopoldmaillard/sceneteract-traces")["test"]
renders = load_dataset("leopoldmaillard/sceneteract-renders")["train"]

meta  = renders.select_columns(["kind", "scene_id", "object_id"])
index = {(r["kind"], r["scene_id"], r["object_id"]): i for i, r in enumerate(meta)}

trace = traces[0]
renders[index[("prompting_view", trace["scene_id"], None)]]["image"]   # task-level input

for step in trace["step_reports"]:                                     # action-level input
    if step["action"] != "navigate_to":
        key = ("target_highlight", trace["scene_id"], step["object_id"])
        renders[index[key]]["image"]

Every scene in the traces has a prompting_view, and every non-navigation (scene_id, object_id) pair has a target_highlight.

Licence and attribution

Released under CC BY-NC 4.0.

If you use this dataset, please cite those works alongside SceneTeract:

• 3D-FRONT — Fu et al., 3D-FRONT: 3D Furnished Rooms with layOuts and semaNTics, ICCV 2021
• 3D-FUTURE — Fu et al., 3D-FUTURE: 3D Furniture Shape with TextURE, IJCV 2021
• PhyScene — Yang et al., PhyScene: Physically Interactable 3D Scene Synthesis for Embodied AI, CVPR 2024

Citation

@misc{maillard2026sceneteract,
      title={{SceneTeract: Probing and Improving Agent-Aware Activity Reasoning in 3D Indoor Scenes}}, 
      author={Léopold Maillard and Francis Engelmann and Tom Durand and Boxiao Pan and Yang You and Leonidas Guibas and Maks Ovsjanikov},
      year={2026},
      eprint={2603.29798},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.29798}, 
}