Learning to Watermark the Denoising Trajectory: Fast and Robust Watermarks for Diffusion Models
cs.CV
Submitted: 2026-03-19
Updated: 2026-11-12
Comments: Accepted in Second Workshop on Technical AI Governance Research (TAIGR) @ ICML 2026
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
The gist: Attributing diffusion-generated images to individual users requires robust recovery of per-image keys, fast verification, and unchanged generator weights.
Terminology
Abstract
Attributing diffusion-generated images to individual users requires robust recovery of per-image keys, fast verification, and unchanged generator weights. To satisfy these requirements, we present DiffMark, which learns to embed arbitrary multi-bit keys along the late denoising trajectory of a frozen generator. We embed the watermark during the final stages of image generation and reinforce it with complementary patterns to improve robustness. The watermark can then be recovered directly from the image, even after spatial transformations, without knowing the key or reversing the generation process. To reduce training memory requirements, we introduce dual-path training: a shortened generation process teaches the encoder to embed the watermark, while images from the full process teach the decoder how to recover it. Theoretically, we bound how much denoising changes a perturbation added during the final sampling steps, give a key-separation condition that we verify on the tested samplers, and identify conditions under which short-path training improves watermark decoding with the full sampler. Empirically, against 10 baselines on three datasets, we show that DiffMark balances robustness, image quality, and extraction efficiency. Across SD2.1, SD3, and FLUX on three datasets, DiffMark achieves 99.4--100% average TPR@1%FPR under distortion, removal, and adversarial attacks, with 89.4 ms extraction, approximately 17x faster than Tree-Ring and Gaussian Shading.
Sources
- TrustMark: Universal Watermarking for Arbitrary Resolution Images
- Classifier-Free Diffusion Guidance
- Latent Consistency Models: Synthesizing High-Resolution Images with Few-Step Inference
- FaceSigns: Semi-Fragile Neural Watermarks for Media Authentication and Countering Deepfakes
- Tree-Ring Watermarks: Fingerprints for Diffusion Images that are Invisible and Robust
- Robust Invisible Video Watermarking with Attention
Related papers
- TEA: Text Encoder Alignment for Robust Concept Erasure in Text-to-Image Models
- Do Egocentric Video-Language Models Capture Both Hand- and Object-Centric Cues?
- Aura: Consistent Multi-Subject Video Generation via VLM-Grounded Semantic Alignment
- PVRF: All-in-one Adverse Weather Removal via Prior-modulated and Velocity-constrained Rectified Flow
- DLWM: Diverse Latent World Models for Efficient Multimodal Reasoning
- Out of Sight, Still in Mind: Token Compression for Omni-LLMs