Research map / Image generation / Autoregressive and latent generation
Your Text Encoder Can Be An Object-Level Watermarking Controller
Naresh Kumar Devulapally, Mingzhen Huang, Vishal Asnani, Shruti Agarwal, Siwei Lyu, Vishnu Suresh Lokhande
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use text-token watermark embedding instead of AquaLoRA UNet fine-tuning (same detector) | AquaLoRA · WikiArt/MS-COCO (48 bits) | PSNR: +7 dB PSNR, >20% bit accuracy | Improved |
| plug watermark token into SD with Textual Inversion personalization (ablation) | Ours + SD · WikiArt (48 bits) | Bit accuracy under DiffPure: 0.88 -> 0.86 | Hurt |
| plug watermark token into styled SD UNet (ablation) | Ours + SD · WikiArt (48 bits) | Bit accuracy under DiffPure: 0.88 -> 0.84 | Hurt |
| choose timestep tau*=8 for training instead of large timesteps (ablation) | Ours · T2I watermarking | image quality vs bit accuracy | Mixed |
| add latent matching loss L_z to the watermark BCE loss (ablation) | Ours · T2I watermarking | imperceptibility | Improved |
| use SAM segmentation masks instead of attention maps for localization (ablation) | Ours · Object-level watermarking | watermark localization accuracy | Improved |
| increase message length to 128 bits (ablation) | Ours · bit-length ablation | bit accuracy: > 91% at 128 bits | No clear effect |
| add object overlap of at least 40% between two watermarked objects (ablation) | Ours · Object-level watermarking | bit accuracy (no attack): 0.94 -> 0.79 | Hurt |
| shrink watermarked object crop from 0.8x to 0.4x size (ablation) | Ours · Object-level watermarking | bit accuracy (no attack): 0.96 -> 0.90 | Hurt |
| increase number of watermarked objects from 1 to 3 (no overlap) (ablation) | Ours · Object-level watermarking | bit accuracy (no attack): 0.99 -> 0.90 | Hurt |
| use style background instead of white background for segmented object (ablation) | Ours · Object-level watermarking | bit accuracy (no attack): 0.99 -> 0.95 | Hurt |
| in-generation token watermark instead of post-generation HiDDeN | HiDDeN · WikiArt | JPEG bit accuracy: 0.88 -> 0.96 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- LaWa — LaWa: Using Latent Space for In-Generation Image Watermarking (ECCV 2024) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- SleeperMark: Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models (CVPR 2025 · similarity 0.90)
- ROBIN: Robust and Invisible Watermarks for Diffusion Models with Adversarial Optimization (NeurIPS 2024 · similarity 0.89)
- PRO: Enabling Precise and Robust Text Watermark for Open-Source LLMs (NeurIPS 2026 · similarity 0.89)
- Watermark-embedded Adversarial Examples for Copyright Protection against Diffusion Models (CVPR 2024 · similarity 0.88)
- RAW: A Robust and Agile Plug-and-Play Watermark Framework for AI-Generated Images with Provable Guarantees (NeurIPS 2024 · similarity 0.88)
- Black-Box Forgery Attacks on Semantic Watermarks for Diffusion Models (CVPR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
