VQA-Illusion/IllusionChar_test

IllusionChar — Test Set

Dataset summary

This repository contains the public test split of IllusionChar, the OCR benchmark introduced in Illusory VQA: Benchmarking and Enhancing Multimodal Models on Visual Illusions. Each metadata row can be paired across source-condition, illusion, filtered-illusion, illusionless-control, and filtered-illusionless-control conditions.

The expected output for an illusion-bearing or source-condition image is an exact, case-sensitive alphanumeric sequence of 3–5 characters. The expected output for an illusionless control is No illusion.

Property	Value
Hugging Face repository	VQA-Illusion/IllusionChar_test
Official split	Test
Task	Illusory OCR / image-to-text / visual question answering
Annotated base examples	3,000
Image variants per example	5
Image format	PNG
Metadata file	df_data.csv
Paper	arXiv:2412.08169
Code	IllusoryVQA/IllusoryVQA

Repository structure

Path	Files	Description	Evaluation target
illusion_images/	3,000	Generated scenes containing an illusory character sequence.	Exact sequence in label.
illusion_images_filtered/	3,000	Filtered versions of the illusion images.	Exact sequence in label.
illusionless_images/	3,000	Matched generated scenes without a hidden sequence.	No illusion.
illusionless_images_filtered/	3,000	Filtered illusionless controls.	No illusion.
raw_images/	3,000	Source-condition sequence images.	Exact sequence in label.
df_data.csv	1	Canonical metadata and sequence targets.	—

All five directories use the same filename stem. For example, illusion_char_0 maps to illusion_char_0.png in each condition.

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 1.5.
label	string	Exact, case-sensitive 3–5 character transcription.

The stored label applies to source-condition, illusion, and filtered-illusion images. For either illusionless directory, override the target with No illusion.

Output encoding

IllusionChar is a sequence-transcription task, not a fixed-class classification task. It therefore has no canonical numeric class-to-label table.

Output type	Valid target
Source-condition, illusion, or filtered illusion	Exact sequence in label
Illusionless or filtered illusionless	No illusion
Sequence length	3, 4, or 5 characters
Character vocabulary	0–9, A–Z, and a–z
Case handling	Case-sensitive

Do not assign one class ID to each complete sequence. The official evaluation treats the outputs as text and reports OCR metrics.

Download

pip install -U huggingface_hub pandas pillow

from huggingface_hub import snapshot_download

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

Command-line alternative:

huggingface-cli download VQA-Illusion/IllusionChar_test \
  --repo-type dataset \
  --local-dir IllusionChar_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/IllusionChar_test",
    repo_type="dataset",
))
df = pd.read_csv(
    root / "df_data.csv",
    dtype={"label": "string"},
    keep_default_na=False,
)

folders = {
    "illusion": "illusion_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():
        target = "No illusion" if condition.startswith("illusionless") else row.label
        records.append({
            "image_name": row.image_name,
            "condition": condition,
            "image_path": root / folder / (row.image_name + ".png"),
            "target_text": target,
        })

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

Evaluation

For source-condition, illusion, and filtered-illusion conditions, use:

• Character error rate (CER) for character-level transcription quality;
• Word error rate (WER) for whole-sequence errors, matching the paper; and
• optional exact-match accuracy for a stricter summary.

Preserve case unless a case-insensitive protocol is explicitly declared. For illusionless conditions, evaluate the exact No illusion response separately or include it as a special text target.

Filtered variants

The paper's released filter pipeline applies Gaussian, averaging, and median blurs followed by grayscale conversion and sharpening. Appendix K contains the exact OpenCV implementation and parameters.

Dataset creation and safety

The authors generated 3–5 character source images, generated English scene descriptions with several language models, and applied the illusion effect with ControlNet. Human reviewers validated quality. The paper reports NSFW screening and exclusion of flagged images from the public release.

The paper reports 3,300 IllusionChar test samples, while the current repository contains 3,000 metadata rows. This card documents the current public repository; use df_data.csv for reproducible indexing.

Important usage notes

• Use df_data.csv as the authoritative index.
• Hugging Face may auto-detect top-level folders as imagefolder classes. These are image conditions, not OCR targets.
• Case is semantically meaningful.
• The correct target for both illusionless conditions is No illusion, irrespective of the stored sequence.
• Keep all variants of an image_name together when creating derived splits.
• The benchmark focuses on OCR under visual illusion rather than open-ended VQA; 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.