Reproduce “VHA: VQA-based Hallucination Assessment for Generative Image Super-Resolution” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Donghun Ryou, Inju Ha, Jaeyoo Park, Bohyung Han
VHA assesses hallucinations in generative super-resolution by asking what evidence is supported in reference and generated images, then comparing answer support in both directions. VHA+ learns to predict VHA scores from image pairs, avoiding repeated vision-language model inference. On VHA-bench, VHA’s Spearman correlation with human ratings is 0.618, versus 0.543 for Hallucination Score prompting. This result is benchmark-specific.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.