ProMark: Proactive Diffusion Watermarking for Causal Attribution

Generative AI (GenAI) is transforming creative work-flows through the capability to synthesize and manipulate images via high-level prompts. Yet creatives are not well supported to receive recognition or reward for the use of their content in GenAI training. To this end, we propose ProMark, a causal attribution technique to attribute a synthetically generated image to its training data concepts like objects, motifs, templates, artists, or styles. The concept information is proactively embedded into the input training images using imperceptible watermarks, and the diffusion models (unconditional or conditional) are trained to retain the corresponding watermarks in generated images. We show that we can embed as many as 2¹⁶ unique water-marks into the training data, and each training image can contain more than one watermark. ProMark can maintain image quality whilst outperforming correlation-based attribution. Finally, several qualitative examples are presented, providing the confidence that the presence of the watermark conveys a causative relationship between training data and synthetic images.

1. Introduction

GenAI is able to create high-fidelity synthetic images spanning diverse concepts, largely due to advances in diffusion models, e.g. DDPM [18], DDIM [23], LDM [28]. GenAI models, particularly diffusion models, have been shown to closely adopt and sometimes directly memorize the style and the content of different training images - defined as “concepts” in the training data [11], [21]. This leads to concerns from creatives whose work has been used to train GenAI. Concerns focus upon the lack of a means for attribution, e.g. recognition or citation, of synthetic images to the training data used to create them and extend even to calls for a compensation mechanism (financial, reputational, or oth-erwise) for GenAI's derivative use of concepts in training images contributed by creatives.