A Framework for Evaluating Zero-Shot Image Generation in Concept-based Explainability

Researchers propose an evaluation framework for using zero-shot text-to-image models to generate synthetic concept datasets for concept-based XAI.

The study tests these generators with similarity, intra-similarity, downstream explanation, and concept-removal metrics.

Results suggest the approach is promising for reducing reliance on large labeled datasets, but faithfulness remains uncertain.

The work highlights both the potential and the current limits of synthetic concept images in explainable computer vision.