A user asks for an image…
Sample several images for the same prompt
Map images to semantic spaceSimilar images land near each other
Split semantic spaceRandom hyperplanes divide the map
Give each region a codeEach line contributes one bit
Sample a region from an appropriate distributionDepending on the group sizes and secret key to preserve the original output distribution
Release one image unchangedPick a random candidate from the winning region
Image generator
golden retriever playing in snow
Generate
image model
semantic encoder
Semantic space · 2D illustration
00
01
10
11

Secret key + group size

001 candidate
011 candidate
102 candidates
Image generator
golden retriever playing in snow
What this gives you

Semantic robustness

original
→
cropped

Survives common content-preserving edits.

Distortion-free

generator
=
with watermark

Same output distribution, averaged over secret keys.

Black-box

Only queries to the generator and semantic encoder.

Can you watermark an image just by choosing it?

Semantic watermarking for generative models

A generated golden retriever running through snow

Distortion-Free Semantic Watermarks
for Black-Box Generative Models

Aviv Yossef, Or Patashnik, Or Zamir

Tel Aviv University

github.com/avivyossef/distortion-free-semantic-watermark

0.0 / 0.0 s