Research map / Image generation / Autoregressive and latent generation
Attributing Image Generative Models using Latent Fingerprints
Guangyu Nie, Changhoon Kim, Yezhou Yang, Yi Ren
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 |
|---|---|---|---|
| Constrain alpha estimate (alpha-regularization) during attribution (ablation) | latent fingerprint attribution, StyleGAN2 FFHQ · FFHQ SG2, no postprocess | attribution accuracy: 0.877 -> 0.983 | Improved |
| Constrain alpha estimate (alpha-regularization) during attribution (ablation) | latent fingerprint attribution, LDM FFHQ · FFHQ LDM, no postprocess | attribution accuracy: 0.364 -> 0.996 | Improved |
| Add Latin hypercube parallel search with multiple initial guesses for alpha (ablation) | latent fingerprint attribution · FFHQ SG2 and LDM, no postprocess | attribution accuracy: SG2 0.711 -> 0.983; LDM 0.872 -> 0.996 | Improved |
| Move fingerprint dimensions from major to minor principal components of Sigma_w (LDM, sigma=1) (ablation) | latent fingerprints, LDM · FFHQ LDM, sigma=1.0 | attribution accuracy: 0.99 -> 0.00 (PC[0:64] to PC[4000:4064]) | Hurt |
| Move fingerprint dimensions from major to minor principal components of Sigma_w (LDM, sigma=1) (ablation) | latent fingerprints, LDM · FFHQ LDM, sigma=1.0 | FID: 33.62 -> 12.77 (lower is better) | Improved |
| Increase fingerprint strength sigma from 1 to 6 on major PCs (SG2) (ablation) | latent fingerprints, StyleGAN2 PC[0:64] · FFHQ SG2 | FID: 129.0 at sigma 0.6 vs 101.3 at sigma 6.0 (mixed direction); accuracy limited benefit beyond small sigma | Hurt |
| Increase key length d_phi from 32 to 128 bits (ablation) | latent fingerprints, StyleGAN2 · FFHQ SG2 | attribution accuracy: 0.982 -> 0.973 (32 to 128 bits) | Hurt |
| Increase key length d_phi from 32 to 128 bits (ablation) | latent fingerprints, StyleGAN2 · FFHQ SG2 | FID: 8.49 -> 9.61 (32 to 128 bits) | Hurt |
| Use latent fingerprints (PC[32:40], 8-bit) instead of shallow fingerprints under Combo postprocess | StyleGAN2 FFHQ · Combo postprocess, unknown attack | attribution accuracy: 0.65 (ours, UK) vs 0.83 (BL, UK); 0.89 vs 0.88 with KN | Mixed |
| Use latent fingerprints (PC[32:40]) instead of shallow fingerprints | StyleGAN2 FFHQ · Combo postprocess | FID: 12.35 vs 100.15 (lower is better) | Improved |
| Use latent fingerprints (PC[32:40]) instead of shallow fingerprints | StyleGAN2 FFHQ · Combo postprocess | SSIM (secrecy): 0.73 vs 0.66 | Improved |
| Postprocess-specific modified LPIPS for attribution (known-attack setting) (ablation) | latent fingerprint attribution · Combo postprocess, KN | attribution accuracy: PC[32:40] UK 0.65 -> KN 0.89 | Improved |
Closest papers by meaning
- ManiFPT: Defining and Analyzing Fingerprints of Generative Models (CVPR 2024 · similarity 0.90)
- WOUAF: Weight Modulation for User Attribution and Fingerprinting in Text-to-Image Diffusion Models (CVPR 2024 · similarity 0.89)
- Scalable Fingerprinting of Large Language Models (NeurIPS 2025 · similarity 0.85)
- OmniMark: Efficient and Scalable Latent Diffusion Model Fingerprinting (AAAI 2025 · similarity 0.85)
- Fingerprinting Denoising Diffusion Probabilistic Models (CVPR 2025 · similarity 0.85)
- How to Trace Latent Generative Model Generated Images without Artificial Watermark? (ICML 2024 · similarity 0.85)
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.
