The paper, a collboration between Denis Lukovnikov, Andreas Müller, Erwin Quiring, and Asja Fischer was accepted at CVPR 2026, and proposes an watermarking approach for autoregressive image generation models to help detect generated images in the wild by embedding an invisible signal deeply into each image.

The starting point from this research was the question “Can we transfer techniques from LLM watermarking, also known as red-green watermarking, into autoregressive image generation”?

The primary challenge in watermarking autoregressive image models is ensuring robust signal extraction. Unlike in large language models, recovering the exact generation tokens in this domain is highly error-prone. The embedded signal is washed out and severely degrades detection accuracy. To overcome this, we introduce two key techniques: clustering the autoregressive token vocabulary for robust watermark color assignment, and adversarially training the latent encoder and tokenizer. Our method achieves SOTA detection robustness with negligible impact on visual fidelity, outperforming concurrent works in this active research area.

Check out the paper (CV open access https://cvpr.thecvf.com/virtual/2026/poster/38427, arXiv https://arxiv.org/abs/2508.06656) and code https://github.com/lukovnikov/ClusterMark.