PACIFIC
GRAPHICS
2026

“Everything should be made as simple as possible, but not simpler.” While this principle of parsimony is commonly applied to design hypotheses in view of observations, it is useful also for the development of computational methods. Simplicity is key for wide adoption in academia as well as use in real world problems. It provides insight into the problem being solved, makes correctness believable, and fosters composition or generalization. I illustrate this with examples in geometric computing, offering some insights from the development of my own work.

Recent advances in multimodal generative models have dramatically expanded our ability to create and reconstruct visual content. Yet generating plausible pixels is only a first step toward models that can truly understand and interact with the world. In this talk, I will discuss our journey from multimodal generation toward unified world modeling, centered on three capabilities: 1) being physical, by representing geometry, material properties, and semantics; 2) being dynamic, by modeling how 3D scenes evolve over space and time; and 3) being actionable, by connecting perception and generation to interaction, control, and embodied simulation. I will present our recent efforts in simulation-ready 3D generation, feed-forward 4D modeling, controllable video generation, and generative embodied simulation, together with new approaches for evaluating world models beyond visual realism. These developments point toward a broader role for graphics: not only synthesizing how the world looks, but building computational representations of how the world is structured, how it changes, and how agents can act within it.

Generative models have transformed our ability to create images, geometry, and other digital content. But designing objects for the physical world requires more than generating something plausible or visually compelling. Physical artifacts must also be manufacturable and satisfy functional and performance requirements, turning design into a search through complex, highly constrained spaces.

It is tempting to think generative models will simply absorb this challenge without the need for explicit optimization. I will argue the opposite: optimization is not merely still relevant; it is becoming more powerful. The recent success of AI is, at its core, a success of large-scale optimization, and it has dramatically expanded what practical optimization can do. But harnessing this power requires rethinking how we represent designs—not treating representations as fixed and building optimizers around them, but instead designing representations for optimization. Through examples from computational design and digital fabrication, I will show how this perspective can make difficult inverse problems tractable and enable new tools for creating objects that work in the physical world.