AI Paper Atlas

Research map / Image generation / Autoregressive and latent generation

The Stable Signature: Rooting Watermarks in Latent Diffusion Models

Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, Teddy Furon

ICCV 2023 · Image generation › Autoregressive and latent generation · approach: Robust diffusion watermarking · 337 citations (OpenAlex)

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 changedWhereMeasuredEffect
fine-tune LDM decoder with a message loss and perceptual lossLDM decoder · MS-COCO text-to-imageFID: -0.3 FIDNo clear effect
merge watermarking into decoder instead of post-hoc HiDDeNText-to-image watermarking · MS-COCO text-to-imagebit accuracy (combined edit): 0.98 -> 0.92Hurt
merge watermarking into decoder instead of post-hoc Dct-DwtText-to-image watermarking · MS-COCO text-to-imagebit accuracy (combined edit): 0.51 -> 0.92Improved
increase perceptual loss weight lambda_i from 0.025 to 0.8 (ablation)Decoder fine-tuning · Combined editbit accuracy: 0.95 -> 0.85Hurt
increase perceptual loss weight lambda_i from 0.025 to 0.8 (ablation)Decoder fine-tuning · Generated imagesPSNR: 24.6 -> 31.4Improved
add JPEG to the attack simulation layer during extractor pre-training (ablation)Watermark extractor · JPEG 50 testbit accuracy: 0.50 -> 0.90Improved
add rotation to the attack simulation layer during extractor pre-training (ablation)Watermark extractor · Rotation 90 testbit accuracy: 0.56 -> 0.99Improved
add crop to the attack simulation layer during extractor pre-training (ablation)Watermark extractor · Crop 0.1 testbit accuracy: 1.00 -> 1.00No clear effect
apply PCA whitening to extractor outputs (ablation)Watermark extractor · Vanilla imagesi.i.d. Bernoulli bit assumptionImproved
attack with the same auto-encoder as the LDMWatermark removal attack · Image-level attackwatermark erasure at quality: PSNR>29 with watermark erasedHurt
increase number of users N for identification (ablation)Identification · Identification at FPR 1e-6accuracy: tau 41 (N=1) -> 44 (N=1000)Hurt
model purification by fine-tuning to remove watermarkNetwork-level attack · Model purificationbit accuracy vs image qualityNo clear effect

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy