Abstract
People frequently capture images in public or semi-public settings (e.g., schools, parks, parties) that include incidental subjects and may share such images with others. The person that captures such images may not have consent from such incidental subjects to publish such images. This disclosure describes generative augmentation techniques that preserve privacy for digital media. Incidental subjects (bystanders) in photographs are automatically protected prior to social sharing. Unlike conventional redaction methods (blurring/pixelation) which degrade visual fidelity, or manual object removal which alters the scene's composition, the techniques semantically substitute unauthorized subjects with context-appropriate, non-identifiable synthetic elements. The synthetic elements can be generated using a latent diffusion model (LDM) guided by a multimodal large language model (MLLM) and depth estimation.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Yakar, Tamar and Labzovsky, Ilia, "Privacy-preserving Generative Inpainting", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10132