Research map / Vision-language and multimodal / Visual question answering

Evaluating Image Hallucination in Text-to-Image Generation with Question-Answering

AAAI 2025 · Vision-language and multimodal › Visual question answering · approach: Visual reasoning diagnostics · 8 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces I-HallA, a VQA-based metric and I-HallA v1.0 benchmark (1,000 multiple-choice QA sets over 200 textbook-derived prompts) for measuring factual image hallucination in text-to-image models, using GPT-4o as the VQA model and human review.

Problem

Text-to-image models often fail to depict factual content accurately, but existing alignment metrics like TIFA only check elements mentioned in the prompt and cannot assess external knowledge or factual visual semantics.

Approach

Build factual and hallucinated image pairs from textbook figures and TTI outputs, use GPT-4o to generate reasoning and multiple-choice factual QA sets, and score TTI images by the fraction of QA sets their images answer correctly under a VQA model.

Key results

• DALL-E 3 achieves the highest I-HallA scores (0.661 science, 0.666 history) among the five TTI models, while factual images score 0.856 and 0.873.
• GPT-4o-based I-HallA scores correlate strongly with 53 human participants (Spearman rho=0.950, Pearson r=0.952, Kendall tau=0.889) on 10 sampled images.
• Using weaker VQA models (InstructBLIP, GPT-4-turbo, GPT-4o-mini) yields lower factual-image scores, indicating GPT-4o is needed for reliable scoring.

Measured effects reported in the paper

Surprising findings

• Prompt refinement with GPT-4o and repeated DALL-E 3 regeneration still failed to produce factually accurate images (e.g., water molecule), suggesting prompt engineering alone does not resolve image hallucination.
• Stable Diffusion models score highest on the Posture (science) and Size (history) compositions despite lower overall image quality.

Limitations

• Relies on GPT-4o as both reasoning generator and VQA judge, and factual images do not score near 1.0 (0.856/0.873), indicating noise or VQA limits.
• Benchmark is limited to science and history textbook content and to five TTI models.
• Human review is needed in the pipeline, and GPT-4o occasionally produces incorrect reasoning.

Experimental setup

Datasets: textbook figures and captions (five textbooks) TTI-generated images from five models

Benchmarks: I-HallA v1.0

Metrics: I-HallA score Spearman correlation with human judgments Pearson correlation Kendall tau

Baselines: TIFA SD v1.4 SD v1.5 SD v2.0 SD XL DALL-E 3

Builds on: TIFA VQ2 Davidsonian Scene Graph

Closest papers by meaning

• PhD: A ChatGPT-Prompted Visual Hallucination Evaluation Dataset (CVPR 2025 · similarity 0.91)
• When Semantics Mislead Vision: Mitigating Large Multimodal Models Hallucinations in Scene Text Spotting and Understanding (NeurIPS 2025 · similarity 0.90)
• The Labeling Problem in Hallucination Detection Benchmarks: An Empirical Evaluation (NeurIPS 2026 · similarity 0.90)
• Fine-Grained Multi Image Object Hallucination Benchmark (CVPR 2026 · similarity 0.89)
• FactTest: Factuality Testing in Large Language Models with Finite-Sample and Distribution-Free Guarantees (ICML 2025 · similarity 0.89)
• Do You Keep an Eye on What I Ask? Mitigating Multimodal Hallucination via Attention-Guided Ensemble Decoding (ICLR 2025 · similarity 0.89)