VQA-Illusion/IllusionAnimals_test

IllusionAnimals — Test Set

Dataset summary

This repository contains the public test split of IllusionAnimals, introduced in Illusory VQA: Benchmarking and Enhancing Multimodal Models on Visual Illusions. Each annotated example is paired across source-condition, illusion, filtered-illusion, illusionless-control, and filtered-illusionless-control conditions.

The animal source-condition images were generated with SDXL-Lightning. English scene descriptions and ControlNet were then used to create the illusory scenes.

Property	Value
Hugging Face repository	VQA-Illusion/IllusionAnimals_test
Official split	Test
Task	Illusion animal classification / visual question answering
Annotated base examples	1,000
Image variants per indexed example	5
Image format	JPEG
Metadata file	df_data.csv
Paper	arXiv:2412.08169
Code	IllusoryVQA/IllusoryVQA

Repository structure

Path	Files	Description	Evaluation target
ill_images/	1,000	Generated images containing the animal illusion.	Animal in label.
illusion_images_filtered/	1,000	Illusion images processed with the paper's filter pipeline.	Animal in label.
illusionless_images/	1,000	Matched generated scenes without an embedded illusion.	No illusion.
illusionless_images_filtered/	1,000	Filtered illusionless controls.	No illusion.
raw_images/	1,001	Source-condition animal images. Exactly 1,000 match metadata IDs; one auxiliary file is not indexed.	Animal in label for indexed files.
df_data.csv	1	Canonical metadata for the 1,000 indexed examples.	—
captions.csv	1	Pool of 1,027 English scene descriptions used in generation.	—

The indexed files share the same stem across directories. For example, IllusionAnimals_1 maps to IllusionAnimals_1.jpg in each condition. The file raw_images/IllusionAnimals_IllusionAnimals.jpg is not referenced by df_data.csv and should be excluded from indexed evaluation.

Metadata schema

Column	Type	Description
image_name	string	Image identifier and shared filename stem.
Pprompt	string	Positive scene prompt used during generation.
Nprompt	string	Negative generation prompt; currently low quality.
illusion_strength	float	Control strength used during illusion generation; currently 2.5.
label	string	Ground-truth animal name.

The label describes the source-condition image and the hidden target in illusion-bearing conditions. For either illusionless directory, override the target with No illusion.

Label mapping

The numeric order below matches the official experiment code. The CSV itself stores animal names.

Numeric ID	Class label	Stored test value
0	cat	cat
1	dog	dog
2	pigeon	pigeon
3	butterfly	butterfly
4	elephant	elephant
5	horse	horse
6	deer	deer
7	snake	snake
8	fish	fish
9	rooster	rooster
10	No illusion	Derived target for the two illusionless directories

Download

pip install -U huggingface_hub pandas pillow

from huggingface_hub import snapshot_download

dataset_dir = snapshot_download(
    repo_id="VQA-Illusion/IllusionAnimals_test",
    repo_type="dataset",
)
print(dataset_dir)

Command-line alternative:

huggingface-cli download VQA-Illusion/IllusionAnimals_test \
  --repo-type dataset \
  --local-dir IllusionAnimals_test

Load all five image conditions

from pathlib import Path
import pandas as pd
from huggingface_hub import snapshot_download

root = Path(snapshot_download(
    repo_id="VQA-Illusion/IllusionAnimals_test",
    repo_type="dataset",
))
df = pd.read_csv(root / "df_data.csv", keep_default_na=False)

label_to_id = {
    "cat": 0,
    "dog": 1,
    "pigeon": 2,
    "butterfly": 3,
    "elephant": 4,
    "horse": 5,
    "deer": 6,
    "snake": 7,
    "fish": 8,
    "rooster": 9,
    "no illusion": 10,
}
folders = {
    "illusion": "ill_images",
    "illusion_filtered": "illusion_images_filtered",
    "illusionless": "illusionless_images",
    "illusionless_filtered": "illusionless_images_filtered",
    "raw": "raw_images",
}

records = []
for row in df.itertuples(index=False):
    for condition, folder in folders.items():
        label_text = "no illusion" if condition.startswith("illusionless") else row.label.casefold()
        records.append({
            "image_name": row.image_name,
            "condition": condition,
            "image_path": root / folder / (row.image_name + ".jpg"),
            "label_id": label_to_id[label_text],
            "label_text": label_text,
        })

evaluation_df = pd.DataFrame(records)
assert evaluation_df["image_path"].map(Path.exists).all()

Filtered variants

The released filtered images are outputs of the preprocessing evaluated in the paper: Gaussian, averaging, and median blurs followed by grayscale conversion and sharpening. Appendix K provides the exact OpenCV implementation and parameters.

Intended use and evaluation

Use this split for animal-illusion classification, constrained-answer VQA, No illusion rejection, and paired robustness comparisons. Recommended metrics are accuracy, macro precision, macro recall, and macro F1. Keep every condition for a given image_name together in any derived split.

Dataset creation and safety

The authors generated animal source-condition images with SDXL-Lightning, generated English scene prompts with several language models, and created the illusion images with ControlNet. Human reviewers validated quality. The paper reports NSFW screening and exclusion of flagged images from the public release.

The paper reports 1,100 IllusionAnimals test samples, while the current repository contains 1,000 metadata rows. This card documents the repository as currently hosted; use df_data.csv for reproducible indexing.

Important usage notes

• Treat df_data.csv as the authoritative index.
• Hugging Face may auto-detect top-level folders as imagefolder classes. These are image conditions, not animal classes.
• For either illusionless condition, the correct answer is No illusion (ID 10), irrespective of the stored animal label.
• Exclude the one unindexed auxiliary file in raw_images/ when reproducing the 1,000-example split.
• The benchmark primarily contains one large hidden category per image; consult the paper for full limitations.

License

This dataset repository declares the MIT license. Users should also review and comply with any applicable terms associated with upstream models and components.

Citation

@misc{rostamkhani2024illusoryvqa,
  title        = {Illusory VQA: Benchmarking and Enhancing Multimodal Models on Visual Illusions},
  author       = {Rostamkhani, Mohammadmostafa and Ansari, Baktash and Sabzevari, Hoorieh and Rahmani, Farzan and Eetemadi, Sauleh},
  year         = {2024},
  eprint       = {2412.08169},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  url          = {https://arxiv.org/abs/2412.08169}
}

Contact

Questions and reproducibility issues can be submitted through the official GitHub repository.