Navigating h-Space for Multi-Attribute Editing in Diffusion Models
J.M. Park, Muhammad Shaheryar, Sanghoon Lee, Soon Ki Jung · 2025
Multi-attribute editing in generative models has been a challenging problem, especially in achieving realistic and disentangled transformations across multiple attributes simultaneously. In this work, we propose an approach for multi-attribute editing in the h-space of diffusion models, where multiple attributes such as aging, gender and eyeglasses can be edited simultaneously. Unlike existing methods that require separate models for each attribute or operate in a highly coupled latent space, our method harnesses the power of a unified framework. We learn interpretable attribute directions in the latent space through supervised training, enabling fine-grained control over specific attributes without affecting others. This disentangled editing allows for complex transformations, such as modifying both age and hairstyle while preserving identity. By performing edits in the h-space, we ensure high-quality, coherent transformations, demonstrating the potential for rich and flexible editing capabilities. The ability to perform multi-attribute modifications in a single, unified model opens up new possibilities for applications in computer vision, digital media, and personalized content creation, making our method a significant advancement in generative modeling.