Diffusion Model Watermarking

Diffusion model watermarking studies how to mark generated images, videos, audio, text, and model behavior so their origins or ownership can later be verified. Common approaches embed keyed signals in initial noise, latent or frequency representations, denoising steps, output pixels, or model weights, then use detectors, inversion, or learned decoders to recover them. Methods increasingly address temporal consistency in video, arbitrary token order in diffusion language models, and semantic binding that can expose forgery or tampering. The central challenge is balancing visual or textual quality and reliable detection against editing, regeneration, removal, and forgery attacks, whose effects depend strongly on the watermark and evaluation protocol.