Research map / Image generation / Autoregressive and latent generation

TRACE: Structure-Aware Character Encoding for Robust and Generalizable Document Watermarking

NeurIPS 2026 · Image generation › Autoregressive and latent generation · approach: Robust diffusion watermarking · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TRACE, a diffusion-based document watermarking method that moves character skeleton keypoints (handle points) via DragDiffusion to embed bits, achieving higher robustness and imperceptibility than font-based and image-based baselines.

Problem

Document watermarking must be robust to cross-media transmission (print-scan, print-camera, screenshot), generalize to unseen fonts, languages and handwritten/artistic styles, and remain imperceptible; existing image-based methods are noise-sensitive and font-based methods rely on pre-defined codebo

Approach

Embed bits by relocating endpoint keypoints of character skeletons with a point-based diffusion editor (DragDiffusion), with heuristic selection of handle/target points (MPE, TPE), automatic mask drawing (MDM), a local consistency loss, and masked region replacement, so no codebook or per-font training is needed.

Key results

• TRACE reaches 100% extraction accuracy on Arial/TNR/Calibri under screenshots at 20-36pt and an average PSNR of 33.34 dB on document images versus 27.19 dB for StegaStamp and 29.60 dB for IHA.
• Under print-camera capture TRACE achieves 92.5-100% accuracy across fonts and sizes, versus roughly 66-85% for AutoStegaFont.
• Ablation shows removing MPE and TPE drops accuracy from 100% to 49.95-68.75%.

Surprising findings

• Removing the stroke-connectivity rule in MPE yields higher PSNR (39.48 dB) but drastically lower accuracy (80.67%), because embedding fails when handle points do not move.

Limitations

• Characters without endpoints (e.g., 'o') cannot carry embedded bits.
• DragDiffusion LoRA fine-tuning and latent optimization add substantial per-document latency.
• Evaluation uses a limited set of fonts, sizes, and a single printer/scanner/camera setup.

Experimental setup

Datasets: Arial, Times New Roman, Calibri text images SimSun, SimHei Chinese characters

Metrics: extraction accuracy PSNR SSIM

Baselines: AutoStegaFont StegaStamp IHA

Builds on: DragDiffusion OpenPose AutoStegaFont CRAFT

Closest papers by meaning

• SleeperMark: Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models (CVPR 2025 · similarity 0.86)
• StableGuard: Towards Unified Copyright Protection and Tamper Localization in Latent Diffusion Models (NeurIPS 2025 · similarity 0.86)
• Break the Trade-off Between Watermark Strength and Speculative Sampling Efficiency for Language Models (ICLR 2026 · similarity 0.85)
• WMAdapter: Adding WaterMark Control to Latent Diffusion Models (ICML 2025 · similarity 0.85)
• Black-Box Forgery Attacks on Semantic Watermarks for Diffusion Models (CVPR 2025 · similarity 0.85)
• Shallow Diffuse: Robust and Invisible Watermarking through Low-Dim Subspaces in Diffusion Models (NeurIPS 2025 · similarity 0.85)