When Pretty Isn't Useful: Investigating Why Modern Text-to-Image Models Fail as Reliable Training Data Generators

arXiv cs.AI · 6 hours ago

Text-to-image diffusion models released between 2022 and 2025 produce visually appealing images but perform worse as synthetic training data for computer vision classifiers despite improvements in visual quality. Classifiers trained exclusively on synthetic data from newer T2I models show consistent accuracy declines on real test data, with accuracy dropping as T2I model generations advance. The findings suggest that generative models' progress in visual fidelity does not translate to generating diverse, representative training data that matches real-world distributions, requiring reconsideration of their utility for synthetic dataset creation.