VHA-bench

VHA-bench is a human-annotated hallucination benchmark for generative image super-resolution. It contains 90 RealSR reference images and outputs from StableSR, DiffBIR, SeeSR, OSEDiff, PURE, and VOSR: 630 images forming 540 GT–SR pairs. The human study contains 11,200 individual ratings from 200 participants. Scores range from 1 to 5; higher values indicate more hallucination relative to the ground truth.

• Paper
• Code
• Project page
• VHA+ checkpoint

Load image pairs

from datasets import load_dataset

dataset = load_dataset("dhryou/VHA-bench")
example = dataset["test"][0]
gt_image = example["gt"]
sr_image = example["sr"]
human_score = example["human_score"]

The test split has 540 rows. Each row includes gt and sr images, sample_id, image_id, method, human_score, and n_ratings.

Files and compatibility with the original benchmark

• test.zip: all 630 original PNG images, unchanged, plus metadata.jsonl connecting the image pairs to their scores.
• mos.csv: the original 540 pair-level mean scores and response counts.
• ratings.csv: the original 11,200 anonymized individual ratings.

For the GitHub evaluation code, extract test.zip into data/benchmark/ and place mos.csv and ratings.csv in the same directory. The archive preserves the original images/GT, images/StableSR, images/DiffBIR, images/SeeSR, images/OSEDiff, images/PURE, and images/VOSR layout. The extra metadata.jsonl is used by Hugging Face and can be left in place.

from huggingface_hub import hf_hub_download
from pathlib import Path
from zipfile import ZipFile

root = Path("data/benchmark")
root.mkdir(parents=True, exist_ok=True)
archive = hf_hub_download("dhryou/VHA-bench", "test.zip", repo_type="dataset")
with ZipFile(archive) as z:
    z.extractall(root)
for filename in ["mos.csv", "ratings.csv"]:
    hf_hub_download("dhryou/VHA-bench", filename, repo_type="dataset", local_dir=root)

The image field in the original CSV files identifies the shared reference image; method selects the corresponding SR subdirectory. human_score is the arithmetic mean of individual ratings, and n_ratings is their count. participant_id in ratings.csv is an anonymized participant identifier.