Keynote Speakers

Prof. Maneesh Agrawala, Stanford University, USA

Title:

Beyond Unpredictable Black Boxes: Designing Generative AI For Iterative Refinement

Abstract:

Human creation of high-quality content often requires significant iteration. People produce a coarse initial draft, and refine it step-by-step towards a final result. While modern generative AI tools are capable of producing surprisingly high-quality content from simple text prompts, they do not support this iterative refinement workflow. Instead today's AI tools are black boxes, making it impossible for users to build a mental/conceptual model that can predict how an input prompt will be transmuted into output content. The lack of predicatability forces users to rely on iterative trial-and-error, repeatedly crafting a prompt, using the AI to generate a result, and then adjusting the prompt to try again. In this talk I'll outline some features generative AI tools should provide to support iterative refinement workflows rather than iterative trial-and-error. These features include hierarchical decomposition of the creation task and consistency of the output content from iteration to iteration. Finally I'll suggest some approaches we might use to build generative AI tools that provide such features and demonstrate a few implementations of these ideas that we have developed in our lab at Stanford.

Speaker's Biography:

Maneesh Agrawala is the Forest Baskett Professor of Computer Science and Director of the Brown Institute for Media Innovation at Stanford University. He works on computer graphics, human computer interaction and visualization. His focus is on investigating how cognitive design principles can be used to improve the effectiveness of audio/visual media. The goals of this work are to discover the design principles and then instantiate them as constraints and controls in generative AI tools. Honors include an Okawa Foundation Research Grant (2006), an Alfred P. Sloan Foundation Fellowship (2007), an NSF CAREER Award (2007), a SIGGRAPH Significant New Researcher Award (2008), a MacArthur Foundation Fellowship (2009), an Allen Distinguished Investigator Award (2014) and induction into the SIGCHI Academy (2021). He was named an ACM Fellow in 2022.

Prof. Eitan Grinspun, University of Toronto, Canada

Title:

Simulate Everything, Everywhere, All At Once

Abstract:

Reduced order modelng (ROM) has long promised the ability to simulate complex dynamics at lightning speed—at the cost of specialization. Traditional ROMs are typically tied to a specific scenario, geometry, and discretization, making them brittle in the face of changing applications, shapes, or numerical methods.

If we could lift these restrictions, then we could unlock the potential of dramatically more expressive ROMS that can accelerate a wide range simulations and applications. These ROMS could absorb data from disparate simulations—even those conducted on different discretizations—and generalize across diverse shapes.

I will present first steps toward these goals. Using neural fields to build continuous, differentiable representations of physical phenomena, we can learn a low-dimensional manifold of kinematic representations not tied to one shape or discretization. These new ROMS enable fast, accurate simulations that can train on—and generalize across—grids, meshes, and point clouds alike, and their generalization across shapes has exciting connections to spectral geometry processing. I believe that these approaches point to a new kind of simulation engine: one that is fast, general, and geometry-aware, bringing us one step closer to simulating everything, everywhere, all at once.

Speaker's Biography:

Eitan Grinspun is Associate Chair, Communications, Mentoring and Inclusion at the Department of Computer Science at the University of Toronto, where he is also a Professor of Computer Science and Mathematics. He was previously the Co-Director of the Columbia Computer Graphics Group at Columbia University (2004-2021), Professeur d'Université Invité at l'Université Pierre et Marie Curie in Paris (2009), a Research Scientist at New York University's Courant Institute (2003-2004), a graduate student at the California Institute of Technology (1997-2003), and an undergraduate in Engineering Science at the University of Toronto (1993-1997). He has been an NVIDIA Fellow (2001), Eberhart Distinguished Lecturer (2003), NSF CAREER Awardee (2007), Alfred P. Sloan Research Fellow (2010-2012), one of Popular Science magazine's "Brilliant Ten Scientists" (2011), and Fast Company magazine's "Most Creative People in Business" (2013). Technologies developed by his lab are used in products such as Adobe Photoshop & Illustrator, at major film studios, and in soft matter physics and engineering research. He has been profiled in The New York Times, Scientific American, New Scientist, and mentioned in Variety. His film credits include The Hobbit, Rise of the Planet of the Apes, and Steven Spielberg's The Adventures of Tintin.

Prof. Minhyuk Sung, KAIST, South Korea

Title:

Inference-time guided generation using diffusion and flow models

Abstract:

Recent breakthroughs in generative AI have transformed the creative process, making it easier than ever to generate realistic images and videos. While the quality of generated outputs has reached unprecedented levels of realism, the challenge now lies in improving controllability and alignment with user preferences. Although text-to-image generative models have become prevalent, text input alone is often insufficient to provide precise spatial or stylistic control. Traditionally, users have provided such inputs through direct interactions, such as mouse clicks, but supporting these traditional input methods has become increasingly challenging. Moreover, despite the widespread use of text-to-image generation, text-image alignment remains far from perfect, particularly for complex prompts. To enhance controllability and alignment with user intent, recent advancements in LLMs have shifted focus beyond scaling training data and model size to scaling inference-time computation, as demonstrated by the AGI-level performance of models like GPT-4o and DeepSeek. In this talk, I will discuss inference-time generation techniques for guided image and video generation, categorized into three main approaches. First, noise manipulation leverages the observation that adjusting noise during the denoising process influences the final output, enabling alignment with user-defined spatial guidance. Second, gradient-descent-based algorithms utilize the expectation of the final output at an intermediate denoising step, which can be interpreted through the lens of score distillation and combined with it. Lastly, particle sampling exploits the stochastic nature of the generative process, branching out the generation to scale up and search for the desired output, albeit at the cost of increased inference computation. I will explore their capabilities, limitations, and future directions.

Speaker's Biography:

Minhyuk Sung is an Associate Professor in the School of Computing at KAIST, affiliated with the Graduate Schools of AI and Metaverse. Previously, he was a Research Scientist at Adobe Research. He earned his Ph.D. from Stanford University under Prof. Leonidas J. Guibas. His research focuses on generating, manipulating, and analyzing various visual data, including images, videos, and 3D data. He has served on program committees for SIGGRAPH Asia (2022-2025), Eurographics (2022, 2024-2025), Pacific Graphics (2023, 2025), ICCV (2025), ICLR (2025), and AAAI (2023, 2024). He received the Asia Graphics Young Researcher Award in 2024.