Your Text Encoder Can Be An Object-Level Watermarking Controller

A learned watermark token conditions image generation; latent-space and attention controls steer watermark placement onto selected objects. — A small text-embedding update controls watermark generation, with attention-based control for selected objects. Figure 2 from the accepted ICCV 2025 paper. Select the figure to enlarge it.

Overview

Watermarking need not apply uniformly to a whole generated image. This work learns a text token that can control a watermark globally or steer it toward selected objects.

How it works

Learn a 768-dimensional watermark-token embedding while keeping the pretrained generation components fixed.
Train the token through an image-to-image generation objective and a pretrained watermark decoder.
Use cross-attention maps and latent controls to target prompt-selected objects during generation.

The paper also examines compatibility with personalized models and textual-inversion concepts.

A reported result

0.99 bit accuracy with a 48-bit payload for the MS-COCO, full-image Stable Diffusion setting without attacks, Table 1.

That is per-bit recovery, not the probability that every message is recovered exactly. Object-level experiments and robustness after image transformations are evaluated separately.

Scope

The clean full-image result does not imply identical recovery for every isolated or cropped object. Watermark placement and recovery depend on attention localization, selected regions, and subsequent transformations.

Source

Accepted paper: Sections 4-5 and Table 1

Bibliographic reference

Naresh Kumar Devulapally, Mingzhen Huang, Vishal Asnani, Shruti Agarwal, Siwei Lyu, Vishnu Suresh Lokhande. (2025). Your Text Encoder Can Be An Object-Level Watermarking Controller.