HalluGen: Synthesizing Realistic and Controllable Hallucinations for Evaluating Image Restoration

TL;DR

HalluGen addresses the challenge of evaluating and mitigating hallucinations in image restoration by introducing a diffusion-based generator that can produce controllable, realistic, and labeled hallucinations. It creates a large, patch-level annotated dataset enabling systematic benchmarking, and demonstrates two key utilities: a hallucination-aware metric (SHAFE) that improves detection over traditional metrics, and a reference-free detector that generalizes to real restoration failures. The work provides a scalable foundation for evaluating and countering safety-critical hallucinations across medical, industrial, and natural images, with open-source release of code and data. While limitations remain in homogeneous regions and explicit semantic control, HalluGen establishes a practical framework for safer deployment of restoration models in critical settings.

Abstract

Generative models are prone to hallucinations: plausible but incorrect structures absent in the ground truth. This issue is problematic in image restoration for safety-critical domains such as medical imaging, industrial inspection, and remote sensing, where such errors undermine reliability and trust. For example, in low-field MRI, widely used in resource-limited settings, restoration models are essential for enhancing low-quality scans, yet hallucinations can lead to serious diagnostic errors. Progress has been hindered by a circular dependency: evaluating hallucinations requires labeled data, yet such labels are costly and subjective. We introduce HalluGen, a diffusion-based framework that synthesizes realistic hallucinations with controllable type, location, and severity, producing perceptually realistic but semantically incorrect outputs (segmentation IoU drops from 0.86 to 0.36). Using HalluGen, we construct the first large-scale hallucination dataset comprising 4,350 annotated images derived from 1,450 brain MR images for low-field enhancement, enabling systematic evaluation of hallucination detection and mitigation. We demonstrate its utility in two applications: (1) benchmarking image quality metrics and developing Semantic Hallucination Assessment via Feature Evaluation (SHAFE), a feature-based metric with soft-attention pooling that improves hallucination sensitivity over traditional metrics; and (2) training reference-free hallucination detectors that generalize to real restoration failures. Together, HalluGen and its open dataset establish the first scalable foundation for evaluating hallucinations in safety-critical image restoration.

Figures (16)

• Figure 1: Circular dependency in hallucination evaluation and our proposed HalluGen solution. Top: Reliable hallucination analysis requires labeled data, but obtaining labels demands expert annotation with high disagreement. Bottom: HalluGen breaks this loop by generating controllable hallucinations with automatic labels, enabling systematic benchmarking, and perceptual studies across domains.
• Figure 2: Existing metrics fail to penalize hallucinations. Across MVTec AD (left) and BraTS (right), PSNR, SSIM, and LPIPS assign higher scores to hallucinated predictions than to slightly blurred but correct images, reflecting a bias toward perceptual sharpness over correctness.
• Figure 3: Representative hallucinations generated by HalluGen across domains. Top: Controlled synthesis of intrinsic and extrinsic hallucinations in our open dataset for MR images. Bottom: Cross-domain generalization to industrial imagery (MVTec AD). HalluGen produces realistic yet semantically incorrect features across domains. More visual results are in the supplementary.
• Figure 4: Controllability of HalluGen. Left: Severity increases with gradient strength $\gamma$. Middle: Severity scales linearly with number of patches while FID stays low. Right: Stable realism across patch sizes (16×16 – 64×64). HalluGen provides fine-grained control over severity, spatial extent, and granularity while preserving realism.
• Figure 5: Visual results of HalluGen on ImageNet. HalluGen generalizes beyond safety-critical domains and can generate realistic hallucinated features as highlighted in red boxes.