PhD Dissertation Proposal by XU Chenshu | Towards User-Controllable Visual Media Generation and Editing

ABOUT THE TALK

Recent generative models have substantially advanced visual content creation, yet practical creative workflows require users to control generated content precisely, intuitively, and predictably. We investigate user-controllable visual media generation and editing across raster images and vector graphics through four progressively forms of control. First, we address identity–style entanglement in personalized image generation, enabling a character’s identity and artistic style to be learned and manipulated independently. Second, we introduce DragNoise, an interactive point-based image editing method that improves editing accuracy, semantic consistency, and computational efficiency. Third, we present DragSVG, which extends drag-based interaction to vector graphics and produces temporally coherent animations that follow user-defined trajectories while preserving the structure and editability of the original SVG. Finally, we propose DragPhysSVG, a framework for scene-aware physical animation that adapts physics-grounded motion to the geometry of a target scene, allowing animated vector objects to interact more naturally with their surrounding environments. Together, these studies demonstrate how suitable representations, interaction mechanisms, and optimization strategies can translate increasingly sophisticated forms of user intent into faithful and predictable visual outcomes, contributing toward generative visual systems that serve as controllable and expressive tools for human creativity.

ABOUT THE SPEAKER

Chenshu XU is a PhD candidate in the School of Computing and Information Systems (SCIS) at Singapore Management University (SMU). She is a member of Visual Understanding and Generation Lab (VUG) @ SMU supervised by Prof. HE Shengfeng. Her research interests lie at the intersection of Computer Vision and Computer Graphics.