Search:
Match:
1 results

Analysis

This paper addresses the challenge of anonymizing facial images generated by text-to-image diffusion models. It introduces a novel 'reverse personalization' framework that allows for direct manipulation of images without relying on text prompts or model fine-tuning. The key contribution is an identity-guided conditioning branch that enables anonymization even for subjects not well-represented in the model's training data, while also allowing for attribute-controllable anonymization. This is a significant advancement over existing methods that often lack control over facial attributes or require extensive training.
Reference

The paper demonstrates a state-of-the-art balance between identity removal, attribute preservation, and image quality.