pix2fact/Pix2FactBenchmark

Pix2Fact: When Vision Is Not Enough — Benchmarking Fine-Grained VQA with Web Verification on High-Resolution Real-World Scenes

🌐 Website: https://fanfan7589.github.io/pix2fact/

📄 Paper: https://arxiv.org/abs/2602.00593

Pix2Fact is a visual question-answering benchmark designed to assess expert-level visual perception and knowledge search. It comprises 1,000 high-resolution (4K+) images spanning eight real-world scenarios, with question–answer pairs meticulously crafted by PhD-holding annotators. Each question requires both fine-grained visual grounding and the integration of external (web) knowledge.

Usage

from datasets import load_dataset

ds = load_dataset("pix2fact/Pix2FactBenchmark", split="train")
print(ds[0]["question"], ds[0]["answer"])
ds[0]["image"]  # PIL.Image

Fields

• image — the high-resolution scene image
• question / answer — the QA pair
• image_url — stable CDN URL of the image (.../resolve/main/images/<file>), also downloadable as a file in the repo
• category — one of the eight scenario categories
• search_query, evidence_*, evidence_url_* — supporting search queries / evidence
• image_description, caption, bounding_box, image_resolution, and other metadata