Explicit Content Generation in Text-to-Image Models: Training Mechanisms, Moderation Failures, and Safety Limits

DOI

https://www.doi.org/10.59256/indjcst.20260502111

Published In

Volume 5, Issue 2 (2026)

Published Online

May-August 2026

Pages

1005-1020

Article Type

Case Study

Check for updates

Text-to-image generative models have rapidly progressed in realism and controllability, enabling wide adoption across creative and commercial domains. However, these systems also generate explicit and harmful imagery in ways that are difficult to prevent reliably. This paper reviews the technical foundations of modern image generators, with emphasis on diffusion-based modeling, latent representations, and multimodal text-image alignment. It then surveys documented moderation ap-proaches and explains why common safeguards fail under paraphrasing, adversarial prompting, long-tail semantics, and distribution shift. Finally, the paper discusses safety limitations that arise from likelihood-based learning objectives and proposes directions for more robust mitigation, including stronger data governance, multimodal safety alignment, evaluation-driven red teaming, and lifecycle monitoring.