Scene-OTC-Bench

Evaluating Scene Critique for Design Intelligence

Scene-Designer: Learning Design Intelligence for Agentic 3D Scene Generation

OmniAI Group of ZJU ACES Lab

Scene-OTC-Bench evaluates whether vision-language models can distinguish better scene designs from poorer ones. Given two or four renderings of the same room, a model must rank the scenes from best to worst, considering physical validity, functional correctness, and layout coherence.

This repository releases the 149-question evaluation benchmark used in Scene-Designer, with 408 original scene images and complete reference rankings. Scene-OTC-15K is the companion training dataset; the release here contains the evaluation set.

Overview: Design Intelligence

We define design intelligence as the ability to both design a scene and critique its quality.

Scene design concerns constructing physically valid, functionally correct, and layout-coherent scenes, while efficiently correcting errors without unnecessary trial and error when problems arise. Scene critique focuses on assessing scene quality along the same dimensions and distinguishing better designs from poorer ones. These two capabilities are closely connected, since effective design requires reliable critique to guide construction, while meaningful critique requires an understanding of what constitutes good scene design.

Scene-OTC-Bench focuses on the critique side of this relationship: can a model recognize which design works better, and consistently order multiple alternatives?

Video Introduction

Watch or download the English introduction · Explore interactive scene demos

The video introduces the broader Scene-Designer project, including design intelligence, data construction, and Critic-guided learning.

Benchmark at a Glance

Task	Questions	Images per question	Total images
Two-scene comparison	94	2	188
Four-scene ranking	55	4	220
Total	149	2 or 4	408

All questions are merged into one test split and one JSON file. The two tasks can be selected using numimages.

The three quality dimensions are:

• Physical validity: whether objects respect physical constraints, including collisions and placement within the room.
• Functional correctness: whether furniture arrangements support their intended use, accessibility, and interaction orientation.
• Layout coherence: whether functionally related objects form coherent groups and the scene has a sensible overall organization.

From Scene Perturbations to Preference Rankings

Scene-OTC constructs comparison examples by perturbing scene layouts. Related scenes share room geometry while differing in furniture placement, orientation, or presence. Perturbations introduce quality degradation, enabling comparisons between better and poorer designs.

The benchmark includes two-scene preference comparisons and the more demanding task of recovering a complete ordering of four scenes. For example, a room can appear open while placing related furniture too far apart, or contain all necessary objects while obstructing a doorway. Evaluating these differences requires reasoning about scene quality beyond object recognition.

The released images and rankings preserve the original evaluation data. Auxiliary images used during annotation are excluded; images/ contains the images supplied to the evaluated models.

Data Format

Scene-OTC-Bench/
├── README.md
├── scene_otc_bench.json
├── SHA256SUMS
├── images/
│   ├── scene_otc_0001_A.png
│   ├── scene_otc_0001_B.png
│   └── ...
└── assets/                  # Dataset-card overview, logo, and video

Each question contains exactly four fields:

{
  "id": "scene_otc_0001",
  "numimages": 2,
  "images": [
    "images/scene_otc_0001_A.png",
    "images/scene_otc_0001_B.png"
  ],
  "ranking_best_to_worst": ["B", "A"]
}

Field	Description
id	Unique question identifier.
numimages	Number of candidate scenes: 2 or 4.
images	Image paths relative to the repository root, ordered A, B, C, D as applicable.
ranking_best_to_worst	Reference scene labels in descending quality order. ["B", "A"] means B is better than A.

The image list specifies the input order; the ranking specifies the answer. Images in assets/ illustrate this dataset card and are not benchmark inputs.

Download and Load

Download the annotations and benchmark images without the presentation assets:

import json
from pathlib import Path
from PIL import Image
from huggingface_hub import snapshot_download

root = Path(snapshot_download(
    repo_id="XuanyuLiu/Scene-OTC-Bench",
    repo_type="dataset",
    allow_patterns=["scene_otc_bench.json", "images/*", "SHA256SUMS"],
))
questions = json.loads((root / "scene_otc_bench.json").read_text())
sample = questions[0]
images = [Image.open(root / path).convert("RGB") for path in sample["images"]]
labels = list("ABCD"[:sample["numimages"]])
print(dict(zip(labels, sample["images"])))
print(sample["ranking_best_to_worst"])

To load the annotation table using Hugging Face Datasets:

from datasets import load_dataset

test = load_dataset("XuanyuLiu/Scene-OTC-Bench", split="test")
pairwise = test.filter(lambda row: row["numimages"] == 2)
four_scene = test.filter(lambda row: row["numimages"] == 4)

The images column stores relative path strings. Loading the annotation table alone does not open the image files; use the downloaded repository root to resolve them, as shown above.

Evaluation

Present candidate images in their listed order, identifying them as A, B, C, and D as applicable. Ask the model to rank the scenes from best to worst based on physical validity, functional correctness, and layout coherence. Keep the reference ranking separate from the model input.

For exact-match accuracy, a prediction is correct only when its complete ordering matches ranking_best_to_worst. Report accuracy separately for the 94 two-scene questions and the 55 four-scene questions. Random complete rankings have expected accuracies of 50% and approximately 4.17%, respectively.

If image order is randomized, remap the reference labels to the new order before scoring. For reproducible comparisons, report the evaluation prompt and any image-order randomization.

Citation

If you use Scene-OTC-Bench, please cite Scene-Designer:

@misc{liu2026scenedesignerlearningdesignintelligence,
      title={Scene-Designer: Learning Design Intelligence for Agentic 3D Scene Generation},
      author={Xuanyu Liu and Hongyan Feng and Miao Pan and Sunlai Chen and Lizhou Liang and Bing Sun and Wenqi Zhang and Xuhong Zhang},
      year={2026},
      eprint={2610.23121},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.23121},
}

Contact

For questions about the benchmark, contact Xuanyu Liu or open a discussion in this dataset repository.