Research map / Image generation / Autoregressive and latent generation
FARI: Robust One-Step Inversion for Watermarking in Diffusion Models
Jindong Yang, Han Fang, Weiming Zhang, Nenghai Yu, Kejiang Chen
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 FARI one-step inversion instead of 50-step DDIM, adversarial distortion | SD v1.5 Tree-Ring · Tree-Ring adversarial | TPR@1e-3: 0.949 -> 0.997 | Improved |
| use FARI instead of one-step DDIM | SD v1.5 Tree-Ring · Tree-Ring adversarial | TPR@1e-3: 0.863 -> 0.997 | Improved |
| use FARI instead of 50-step DDIM | SD v2.1 Gaussian Shading · Gaussian Shading adversarial | bit accuracy: 0.9755 -> 0.9824 | Improved |
| use FARI instead of naive inversion | SD v3.5M Tree-Ring · Tree-Ring adversarial | TPR@1e-3: 0.161 -> 0.978 | Improved |
| use FARI instead of one-step DDIM | SDXL-Turbo Tree-Ring · Tree-Ring adversarial | TPR@1e-3: 0.201 -> 0.855 | Improved |
| use ExactDPM instead of 50-step DDIM | SD v2.1 Tree-Ring · Tree-Ring adversarial | TPR@1e-3: 0.962 -> 0.906 | Hurt |
| use BELM instead of 50-step DDIM | SD v1.5 Tree-Ring · Tree-Ring adversarial | TPR@1e-3: 0.949 -> 0.592 | Hurt |
| add adversarial fine-tuning to AMED-Solver (AMED-dagger) vs FARI | AMED-Solver · SD v1.5 Gaussian Shading adversarial | bit accuracy: 0.9656 -> 0.9834 | Hurt |
| train FARI with adversarial distortions (vs without) (ablation) | FARI · Image reconstruction | PSNR: 14.60 -> 10.61 | Hurt |
| increase LoRA rank above 1 (ablation) | FARI · LoRA rank ablation | performance | No clear effect |
| train with multi-step inversion (NFE>1) instead of one step (ablation) | FARI · NFE ablation | performance | Hurt |
| train with all noise types instead of none, testing unseen distortions (ablation) | FARI · Noise-type generalization | MSE of reconstructed noise | Improved |
Closest papers by meaning
- CoSDA: Enhancing the Robustness of Inversion-based Generative Image Watermarking Framework (AAAI 2025 · similarity 0.92)
- PGID: Progressive Guided Inversion and Denoising for Robust Watermark Detection (NeurIPS 2026 · similarity 0.90)
- Shallow Diffuse: Robust and Invisible Watermarking through Low-Dim Subspaces in Diffusion Models (NeurIPS 2025 · similarity 0.89)
- T2SMark: Balancing Robustness and Diversity in Noise-as-Watermark for Diffusion Models (NeurIPS 2025 · similarity 0.89)
- GaussMarker: Robust Dual-Domain Watermark for Diffusion Models (ICML 2025 · similarity 0.88)
- ROBIN: Robust and Invisible Watermarks for Diffusion Models with Adversarial Optimization (NeurIPS 2024 · 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.
