Research map / Image generation / Autoregressive and latent generation
BiSLW: Bi-Spectral Latent Watermarking for Generative Diffusion Models
Aryan Pandit
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 |
|---|---|---|---|
| Embed watermark in both low- and high-frequency latent bands vs low only (ablation) | BiSLW · AI-generated images 48-bit | Bit acc. (combined attack): 0.88 -> 0.93 | Improved |
| Embed in both bands vs high only (ablation) | BiSLW · AI-generated images 48-bit | Bit acc. (combined attack): 0.90 -> 0.93 | Improved |
| Embed in both bands vs low only, on image quality (ablation) | BiSLW · AI-generated images 48-bit | PSNR: 38.96 -> 37.71 | Hurt |
| Increase latent fidelity weight lambda_z from 0.5 to 20 (ablation) | BiSLW · AI-generated images 48-bit | PSNR / combined accuracy: PSNR 36.85 -> 39.59; Comb acc 0.96 -> 0.88 | Mixed |
| Use spectral radius r=0.25 instead of 0.15 or 0.35 (ablation) | BiSLW · AI-generated images 48-bit | Bit acc. (combined attack): 0.89 / 0.90 -> 0.93 | Improved |
| Increase embedding strength (alpha_L, alpha_H) from (0.6,0.2) to (1.0,0.5) (ablation) | BiSLW · AI-generated images | PSNR / combined accuracy: PSNR 38.20 -> 36.95; Comb acc 0.89 -> 0.95 | Mixed |
| Use average fusion instead of learned weighted fusion (ablation) | BiSLW decoder · AI-generated images 48-bit | Bit acc. (combined attack): 0.93 -> 0.93 | No clear effect |
| Use average fusion instead of max fusion (ablation) | BiSLW decoder · AI-generated images 48-bit | Bit acc. (combined attack): 0.92 -> 0.93 | Improved |
| Increase payload from 32 to 128 bits (ablation) | BiSLW · AI-generated images | PSNR / combined accuracy: PSNR 38.82 -> 36.34; Comb acc 0.98 -> 0.87 | Hurt |
| Replace LaWa with BiSLW | LaWa* · AI-generated images 32-bit | PSNR: 34.25 -> 37.40 | Improved |
| Replace LaWa with BiSLW against diffusion regeneration t*=250 | LaWa* · Regeneration attack | Bit accuracy: 0.94 -> 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
- Harnessing Frequency Spectrum Insights for Image Copyright Protection Against Diffusion Models (CVPR 2025 · similarity 0.89)
- Shallow Diffuse: Robust and Invisible Watermarking through Low-Dim Subspaces in Diffusion Models (NeurIPS 2025 · similarity 0.88)
- Black-Box Forgery Attacks on Semantic Watermarks for Diffusion Models (CVPR 2025 · similarity 0.88)
- WMAdapter: Adding WaterMark Control to Latent Diffusion Models (ICML 2025 · similarity 0.87)
- Gaussian Shading: Provable Performance-Lossless Image Watermarking for Diffusion Models (CVPR 2024 · similarity 0.87)
- T2SMark: Balancing Robustness and Diversity in Noise-as-Watermark for Diffusion Models (NeurIPS 2025 · similarity 0.87)
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.
