VI-Probe

This dataset accompanies the paper Do VLMs Perceive or Recall? Probing Visual Perception vs. Memory with Classic Visual Illusions.

VI-Probe is a controllable visual-illusion benchmark for probing whether vision-language models respond based on visual perception or recall/memorized priors. The dataset includes classic visual illusion cases with original images, graded perturbations, matched controls, visual guide variants, and reverse-prompt variants.

This folder is organized by illusion case for easy inspection, with flat metadata for Hugging Face preview and grouped reporting.

Files

• test.parquet: viewer-ready table with an image column and one row per image-prompt pair, including reverse-prompt rows (is_rp=1).
• metadata.csv: CSV copy of the same metadata for lightweight scripting.
• summary_by_type.csv: counts grouped by category, type, reverse-prompt flag, and answer.
• summary_by_case_type.csv: counts grouped by class and type.
• dataset_order.csv: 27-class order used in the benchmark.
• class_image_type_counts.csv: image counts for each class and sample type.

Folder Structure

The image files are organized by illusion category and case for manual inspection:

VI-Probe/
  README.md
  test.parquet
  metadata.csv
  dataset_order.csv
  summary_by_type.csv
  summary_by_case_type.csv
  class_image_type_counts.csv
  1_illusion_size/
    1_MullerLyerIllusion/
      original/
      perturbed/
      original_control/
      perturbed_control/
      original_with_guide/
      perturbed_with_guide/
    ...
  2_illusion_color/
    12_CornswweetIllusion/
      original/
      perturbed/
      original_control/
      perturbed_control/
      original_with_guide/
      perturbed_with_guide/
    ...
  3_illusion_orientation/
    19_HeringIllusion/
      original/
      perturbed/
      original_control/
      perturbed_control/
      original_with_guide/
      perturbed_with_guide/
    ...

Each row in test.parquet and metadata.csv points to one image-prompt pair. The same image appears twice when both the regular prompt and reverse prompt are included; use is_rp to distinguish them.

Loading

Load the benchmark split directly with Hugging Face Datasets:

from datasets import load_dataset

ds = load_dataset("xxsun/VI-Probe", split="test")

example = ds[0]
image = example["image"]
prompt = example["prompt"]
answer = example["answer"]

Useful columns:

column	description
image	PIL image object loaded by Hugging Face Datasets
image_path	relative path to the image file in the repository
case_id	1-27 illusion case id
illusion_name	illusion case name
category	size, color, or orientation
type	original, perturbed, original_control, perturbed_control, original_with_guide, or perturbed_with_guide
base_scale	base stimulus scale
strength	perturbation strength
is_rp	whether the row uses the reverse prompt
prompt	VQA prompt
answer	Yes or No
answer_value	numeric answer, 1 for Yes and 0 for No

Example grouping for reporting:

df = ds.to_pandas()
summary = (
    df.groupby(["type", "is_rp", "category"])["answer"]
    .value_counts()
    .reset_index(name="count")
)

Overall Counts

• Unique images: 9570
• Metadata rows including reverse prompts: 19140
• Split: test
• Non-reverse prompt rows: 9570
• Reverse prompt rows: 9570

Non-RP Counts By Type

type	count
original	290
perturbed	2900
original_control	290
perturbed_control	2900
original_with_guide	290
perturbed_with_guide	2900

Non-RP Counts By Category

category	count
size	3993
color	2310
orientation	3267

Reporting Recommendation

Use the flat metadata.csv for evaluation and group results by type, is_rp, category, and case_id. Keep the image files organized by case so per-illusion inspection remains easy.

Citation

If you use VI-Probe, please cite:

@inproceedings{sun2026vlms,
  title={Do VLMs Perceive or Recall? Probing Visual Perception vs. Memory with Classic Visual Illusions},
  author={Sun, Xiaoxiao and Li, Mingyang and Yuan, Kun and Sun, Min Woo and Endo, Mark and Wu, Shengguang and Li, Changlin and Zhang, Yuhui and Wang, Zeyu and Yeung-Levy, Serena},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={25861--25870},
  year={2026}
}