HC-Bench

HC-Bench is a compact multi-part image benchmark for evaluating recognition and prompting robustness, especially in hidden-content scenes. It contains:

• object/ — 56 base images and 56 hidden variants of the same lemmas, plus prompts and metadata.
• text/ — 56 Latin/English and 56 Chinese lemma–description pairs with matching PNGs.
• wild/ — 53 in-the-wild images for additional generalization checks.

---

Repository structure

HC-Bench/
├─ object/
│  ├─ base/                     # 56 base images (7 types × 8 lemmas)
│  ├─ hidden/                   # 56 hidden-content variants (same lemmas)
│  ├─ image\_base.txt            # 7 types and their 8 lemmas each
│  ├─ image\_generate\_prompts.txt# per-lemma scene prompts used for generation
│  └─ lemmas\_descriptions.json  # \[{Type, Lemma, Description}] × 56
├─ text/
│  ├─ Latin/                    # 28 English PNGs
│  ├─ Chinese/                  # 28 Chinese PNGs
│  ├─ English\_text.json         # 56 entries (Type, Length, Rarity, Lemma, Description)
│  └─ Chinese\_text.json         # 56 entries (Type, Length, Rarity, Lemma, Description)
└─ wild/                        # 53 PNGs

---

Contents

object/

• base/: Canonical image per lemma (e.g., Apple.jpg, Einstein.png).
• hidden/: Composite/camouflaged image for the same lemma set (e.g., apple.png, einstein.png).
• image_base.txt: The 7 high-level types and their 8 lemmas each (Humans, Species, Buildings, Cartoon, Furniture, Transports, Food).
• image_generate_prompts.txt: Per-lemma prompts used to compose/generate scenes (e.g., “A monorail cutting through a futuristic city with elevated walkways” for notredame).
• lemmas_descriptions.json: Minimal metadata with {Type, Lemma, Description} aligned 1:1 with the 56 lemmas.

text/

• Latin/ & Chinese/: 28 images each (total 56).
• English_text.json & Chinese_text.json: 56-entry lists pairing lemmas to descriptions in two languages.
  (Note: The English_text.json/Chinese_text.json files include extra fields Length and Rarity for flexibility.)

wild/

• 53 natural/urban scenes for robustness and transfer evaluation.

---

Quick start (🤗 Datasets)

HC-Bench uses the ImageFolder/“imagefolder” style. Class labels are inferred from directory names when present (e.g., base, hidden). If you prefer raw images without labels, pass drop_labels=True.

Load object/base and object/hidden

from datasets import load_dataset

base = load_dataset(
    "imagefolder",
    data_files="https://huggingface.co/datasets/JohnnyZeppelin/HC-Bench/resolve/main/object/base/*",
    split="train",
    drop_labels=True,  # drop automatic label inference
)

hidden = load_dataset(
    "imagefolder",
    data_files="https://huggingface.co/datasets/JohnnyZeppelin/HC-Bench/resolve/main/object/hidden/*",
    split="train",
    drop_labels=True,
)

Load wild/

wild = load_dataset(
    "imagefolder",
    data_files="https://huggingface.co/datasets/JohnnyZeppelin/HC-Bench/resolve/main/wild/*",
    split="train",
    drop_labels=True,
)

Load the JSON metadata (English/Chinese)

from datasets import load_dataset

en = load_dataset(
    "json",
    data_files="https://huggingface.co/datasets/JohnnyZeppelin/HC-Bench/resolve/main/text/English_text.json",
    split="train",
)
zh = load_dataset(
    "json",
    data_files="https://huggingface.co/datasets/JohnnyZeppelin/HC-Bench/resolve/main/text/Chinese_text.json",
    split="train",
)

Docs reference: load_dataset for JSON & files, and ImageFolder for image datasets.

---

Pairing base/hidden with metadata

Filenames differ in casing/spaces between base/ (Apple.jpg) and hidden/ (apple.png). Use object/lemmas_descriptions.json as the canonical list of 56 lemmas and join by Lemma:

import pandas as pd
from datasets import load_dataset

# 1) Canonical lemma list
lemmas = load_dataset(
    "json",
    data_files="https://huggingface.co/datasets/JohnnyZeppelin/HC-Bench/resolve/main/object/lemmas_descriptions.json",
    split="train",
).to_pandas()

# 2) Build (lemma -> file) maps
def to_lemma(name):  # normalize filenames to lemma
    import re, os
    stem = os.path.splitext(os.path.basename(name))[0]
    return re.sub(r"\s+", "", stem).lower()

base_ds = load_dataset(
    "imagefolder",
    data_files="https://huggingface.co/datasets/JohnnyZeppelin/HC-Bench/resolve/main/object/base/*",
    split="train",
    drop_labels=True,
)
hidden_ds = load_dataset(
    "imagefolder",
    data_files="https://huggingface.co/datasets/JohnnyZeppelin/HC-Bench/resolve/main/object/hidden/*",
    split="train",
    drop_labels=True,
)

import os
base_map  = {to_lemma(x["image"].filename): x["image"] for x in base_ds}
hidden_map= {to_lemma(x["image"].filename): x["image"] for x in hidden_ds}

# 3) Join
lemmas["base_image"]   = lemmas["Lemma"].apply(lambda L: base_map.get(L.lower()))
lemmas["hidden_image"] = lemmas["Lemma"].apply(lambda L: hidden_map.get(L.lower()))

---

---

Statistics

• object/base: 56 images
• object/hidden: 56 images
• text/Latin: 28 images
• text/Chinese: 28 images
• wild: 53 images

---

Citation

If you use HC-Bench, please cite:

@misc{li2025semvinkadvancingvlmssemantic,
      title={SemVink: Advancing VLMs' Semantic Understanding of Optical Illusions via Visual Global Thinking}, 
      author={Sifan Li and Yujun Cai and Yiwei Wang},
      year={2025},
      eprint={2506.02803},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2506.02803}, 
}

---