There are no visible watermarks, but model makers can use steganographic codes to identify outputs from their own models.

Text-to-Image Models Leave Identifiable Signatures: Implications for Leaderboard Security

https://arxiv.org/pdf/2510.06525