Nanyang Technological University, Singapore

Talk title and abstract

Beyond Prompting: Creative Control in Visual Generation

Human-AI co-creation requires generative models that do more than produce visually plausible outputs from one-shot prompts. Creative users often need precise, iterative control over who appears, what should be changed or preserved, where new content should be placed, and how the final image should look. In this talk, I will present four recent works that explore visual generation through this lens of controllability. We begin with video-reference talking avatar generation, where identity and expression cues are drawn from rich temporal references. We then discuss precise object and effect removal, highlighting the importance of preserving user-specified scene content. Next, we examine 3D-aware object insertion as a form of spatial and design-oriented co-creation. Finally, we present defocus blur control in text-to-image diffusion, showing how photographic parameters can become intuitive creative handles. Together, these works suggest a path from passive generation toward controllable visual systems that better support human intent, authorship, and refinement.

Chen Change (Cavan) Loy is Tan Lip-Bu Professor in Artificial Intelligence at the College of Computing and Data Science, Nanyang Technological University, Singapore. He is the Director of MMLab@NTU and Co-Associate Director of S-Lab. His research interests include large multimodal models, generative AI, spatial intelligence, representation learning, and visual content creation. Cavan currently serves or has served as an Associate Editor for leading journals including IJCV, TPAMI, and CVIU. He has also served as an Area Chair or Senior Area Chair for major conferences such as CVPR, ICCV, ECCV, ICLR, and NeurIPS. He was Program Co-Chair for CVPR 2026 and will serve as General Co-Chair for ACCV 2028.

Canva

Talk title and abstract

Reimagining Design Intelligence: The Layer Is Where Co-Creation Begins

A flat image is where co-creation ends; a layered design is where it begins. Today's generative models produce stunning pictures — but a picture of a design is not a design. Real designs are layered, structured, and editable; real creation is iterative, with intent that evolves as people make. Co-creation therefore demands outputs humans can take over at any moment: representations that preserve authorship instead of locking people out of their own work. In this talk, I will share our team's journey toward design-native intelligence along three threads. Generation: ART introduces variable multi-layer transparent image generation, MRT unifies text-to-layer, image-to-layer, and layer-to-layer generation in a single model, and PixelART shows that image-to-layer decomposition can be learned directly in pixel space — with neither a VAE nor T2I pretraining. Understanding: DAD (Detect Anything in Graphic Design) parses the world's existing flat designs back into editable elements, letting AI meet creators where their content already lives. Restyling: ReChart transfers visual style across charts while preserving the underlying data and layout, treating editing as a first-class generative task. Several of these capabilities already ship in products used by millions of creators. Together, they converge on design-native foundation models that generate, understand, and edit designs in their native layered form — turning generative AI from a picture generator into a genuine co-creator of living design documents.

Yuhui (Ryan) Yuan is a Research Director at Canva CORE and the founder of Canva Research Lab in China, where he leads a research team focused on building next-generation graphic design foundation models. His current work centers on frontier multimodal generation, multi-layer visual content generation, and graphic design editing models, with the goal of advancing AI systems that can transform how people create and edit visual content.

Before joining Canva, he spent eight years at Microsoft as a Senior Researcher, working on cutting-edge computer vision problems including semantic segmentation, object detection, scene understanding, document intelligence, and generative AI applications. Several of his research contributions have been integrated into Microsoft products, including Azure Form Recognizer and Microsoft Designer.

Ryan received his Ph.D. in Computer Vision from the University of Chinese Academy of Sciences, where he worked on semantic segmentation, and his M.S. in Computer Science from Peking University. His representative research includes works such as OCNet, OCRNet, HDETR, Glyph-ByT5, SPO, and ART, as well as publications in top venues including ICCV, ECCV, IJCV, and ICLR.

University of Modena and Reggio Emilia, Italy

Talk title and abstract

A Question of Control

Generative AI has transformed image creation, but meaningful artistic expression requires more than high-quality synthesis: it requires control. While controllable image generation is an active area of research, different artists demand different forms of control depending on their creative goals and workflows. This talk examines AI co-creation from a computer graphics perspective, where interactive authoring and user control have long been central themes. Through examples from artistic practice, I will discuss why control is essential for creative expression, how existing approaches address different aspects of the problem, and what challenges remain. Finally, I will briefly explore how these ideas extend beyond 2D image generation to 3D content creation, highlighting opportunities for bringing the principles of AI-assisted co-creation to three-dimensional design.

Fabio Pellacini is a Full Professor of Computer Science at the University of Modena and Reggio Emilia, Italy. His research focuses on the use of computer graphics methods to address design problems, with particular emphasis on creative, design, and entertainment industries. His work combines algorithms, efficient systems, numerical methods, and machine learning to support intuitive and interactive creation of complex 3D scenes, enabling both professional designers and novices to work with significantly less effort.

Before joining the University of Modena and Reggio Emilia, Italy, he was an Associate and Full Professor at Sapienza University of Rome, an Assistant and Associate Professor at Dartmouth College, a Visiting Assistant Professor at Cornell University, and a researcher in the R&D division of Pixar Animation Studios. He received his M.S. and Ph.D. in Computer Science from Cornell University and a Laurea degree in Physics from the University of Parma.

Pellacini has made significant contributions to appearance design, material and lighting editing, appearance fabrication, visualization of design workflows, and cloud-based collaborative design. He has received a National Science Foundation CAREER Award and an Alfred P. Sloan Fellowship, and regularly publishes in leading computer graphics venues.

Adobe Research

Talk title and abstract

Title coming soon

Abstract coming soon.

Mira is senior principal scientist at Adobe leading research in Human Computer Interaction (HCI) at the intersection of video interaction and AI agents. Mira leads the STORIE Lab, which focuses on research on storytelling and interactive experiences with generative AI. Mira and her team build new tools that make video and audio creation easier, more fun, and more accessible to a wider audience. She is passionate about multimodal interaction, AI agents, and experiences at the intersection of the physical and digital world. She finished her Ph.D. in computer science at the University of Washington with David Salesin, Michael Cohen and Steven Drucker. Her thesis focused on novel interaction techniques for collecting and organizing web content. She was an undergraduate at the University of Michigan in Ann Arbor and completed her B.S.E. in Computer Engineering.

Mira helped write and edit: No Code Required: Giving Users Tools to Transform the Web, was featured on Forbes.com and helped lead Project Blink. She has presented her research on stage at Adobe Summit and Adobe MAX.