Tel Aviv's NoisePrints Proves AI Image Ownership Without Watermarks or Model Access
A new scheme uses the random seed behind a diffusion model's initial noise as a cryptographic proof of authorship, no model weights required.
Takeaways
- NoisePrints uses the diffusion model's random seed as a distortion-free watermark, requiring no changes to generation. Paper
- Verification needs only the seed and output, not model weights, enabling third-party authorship checks on private models.
- A hash-based noise sampler makes seed recovery from generated content computationally infeasible.
- Zero-knowledge proofs let creators prove ownership without ever revealing the seed itself.
- Robust to Gaussian blur, noise, and brightness attacks across SD 2.0, SDXL, FLUX Schnell, and video models.
- Reference code and pipelines are available at github.com/nirgoren/NoisePrints.
Proving who generated an AI image is quickly becoming a business problem, not just an academic one. When a studio, a stock provider, or an artist wants to demonstrate that a specific piece of visual content came from their private diffusion model, they usually have two bad options: embed a watermark that visibly degrades quality, or run heavy re-inversion procedures that require full access to the model weights. A new paper called NoisePrints takes a different route, treating the random seed used to start the diffusion process as the watermark itself.
The authors, from Tel Aviv University and collaborators, describe the approach as distortion-free, lightweight, and verifiable by a third party who never sees the model. They validate it on several state-of-the-art diffusion systems for images and video, showing that verification needs only the seed and the output.
Why existing watermarks keep failing
Watermarking diffusion outputs has largely followed two families. Post-hoc pixel-space marks are fragile under compression or filtering. Methods like Tree-Ring manipulate the initial noise pattern in Fourier space, then invert the diffusion process to recover it. Inversion-based schemes are computationally expensive, and existing implementations require access to model weights, which makes them impractical at scale.