The field of text-to-image models is moving towards improved safety, precision, and control. Recent developments focus on mitigating semantic leakage, detecting and mitigating implicit malicious intentions, and preserving identity in generated images. Noteworthy papers include DeLeaker, which introduces a dynamic inference-time approach to mitigate semantic leakage, and NDM, which proposes a noise-driven detection and mitigation framework against implicit sexual intentions. Other notable works include SELECT, a dynamic anchor selection framework for precise concept erasure, and Patronus, a defensive framework that safeguards text-to-image models against white-box adversaries.