Presentation Information
[1Yin-B-26]Continuous Attribute Control in Deep Image Editing Models using Attention
〇Koya Mitomi1, Takashi Matsubara1,2 (1. Hokkaido University, 2. AI Lab, CyberAgent, Inc.)
Keywords:
Deep Learning,Image Editing
Diffusion models have driven the rapid spread of natural language–based image editing. However, continuously controlling specific object attributes such as color, shape, and texture in images remains a challenging task. In this study, we address this issue by focusing on Multi-Modal Attention layers. These layers enable interactions between textual and visual representations. We propose a method to control how strongly specific text tokens affect image tokens in these layers. Experimental results across a variety of image editing tasks demonstrate that the proposed method enables smooth and natural variations of attributes.