Beyond CAD

Adrien and Changjian will introduce the background of this workshop and their collaboration stories and outcomes.

In this talk, I will describe our recent efforts in digitizing, representing, and interacting with complex motion in 3D. Particularly, I will cover two complementary directions. First, in the capturing direction, I will introduce ActionMesh, an efficient feedforward model that transforms standard monocular videos into fully 4D animated meshes. ActionMesh not only animates meshes across time but also establishes temporal correspondence, with the entire pipeline running in just a few minutes. Second, in the anticipating (motion) direction, I will discuss how we can best observe and frame these captured motions. I will present -- LAMP -- a novel, specialized language, heavily inspired by traditional cinematographic conventions, that allows users to author camera behaviors that intelligently follow and anticipate scene dynamics and character movement. I will conclude with personal reflections of our attempts to unify capturing and editing of 4D motion towards creative digital storytelling. More information can be found at https://geometry.cs.ucl.ac.uk/.

Flow modeling is fundamental in fluid simulation, and recently constitutes the basis for powerful generative modeling. In this talk, we investigate problems with flow modeling for both physical fluid simulation and generative modeling. In particular, we show that for PINN-based fluid solvers, the strong form NS equations for flow modeling cannot deliver accurate results for complex 3D domains, and propose to fix the issue with weak-form multi-scale PINN (MUSA-PINN). For generative modeling, we show that standard flow matching model under samples for highly combinatorial data spaces, and propose to fix the issue with combinatorial stochasticity, which enables high quality generation for sparse and structured data.

Recent advances in world models for vision have been driven largely by powerful 2D generative models that predict future scene states from visual observations and planned actions, often with little or no built-in knowledge of 3D geometry or physical dynamics. In this talk, I will discuss why physical inductive biased, such as explicit 3D structure, remain essential for world modelling. I will present recent work showing how incorporating geometric and physical priors into learned models leads to more generalisable behaviour, significantly more efficient inference, and better control over predictions.

Which representation of 3D geometry is the best: meshes, point clouds, implicit functions, or something else? Savvy graphics practitioners know that there is no one correct answer to this question; it depends on what one wants to do with that geometry. When it comes to creating and manipulating 3D shapes and scenes, different creative goals may benefit from different ways of *abstracting* 3D geometry---eliding irrelevant details and focusing on aspects that matter for the task at hand. In this talk, I'll discuss some recent efforts to build systems which can synthesize such abstractions from high-level descriptions of user goals. I'll also compare this approach to the currently-dominant paradigm of "3D generative AI" and argue why it is a better way forward for making the practice of 3D modeling both easier and more fulfilling.

As inference-based generative models reshape design, it is easy to imagine optimization fading into the background. This talk argues the opposite. In engineering design, generating plausible solutions is not enough. Artifacts must satisfy stringent fabrication and performance requirements, which demands systematic search through complex design spaces. But 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. This talk asks how we can harness these advances, together with novel inference techniques, to revisit longstanding challenges in computational design.

Boundary representations are central to modern CAD systems, yet they remain challenging for learning-based methods due to the tight coupling of continuous geometry and discrete topology. This difficulty spans simple analytic primitives, such as planes and cylinders described by a few parameters, to complex freeform B-Spline surfaces with intricate structure. This talk presents two representation strategies for robust B-Rep reconstruction and generation, designed to encode geometric and topological structure beyond direct heterogeneous graph prediction. Building on these foundations, it further explores language-guided B-Rep editing, where high-level semantic instructions can be injected and aligned with a pretrained geometric latent space, enabling semantic CAD editing without relying on construction history.

Generating high-quality 3D assets from minimal user input has long been a fundamental challenge in computer graphics and computer vision. Over the past few years, the field has undergone a rapid transformation—from leveraging powerful 2D generative models for 3D synthesis to developing native 3D generative frameworks that directly model geometry and structure. In this talk, I will trace this evolution through three works that address key challenges in consistency, efficiency, and representation. I begin with SweetDreamer, which tackles the longstanding Janus problem in text-to-3D generation by introducing geometry-aware mechanisms that better align 2D diffusion priors with underlying 3D structure. Next, I present CraftsMan, a native 3D generative framework that replaces costly optimization-based pipelines with a Diffusion Transformer (DiT), enabling coarse geometry generation in seconds while preserving fine-grained surface details through a dedicated refinement stage. Finally, I discuss LoST (Level-of-Semantics Tokens), a hierarchical tokenization framework that represents shapes as semantically structured sequences, enabling scalable autoregressive modeling of complex 3D content. Together, these works illustrate the field's transition from 2D-guided 3D generation toward native 3D foundation models. I will conclude with a discussion of emerging directions, including multimodal generative models that jointly reason about language, images, videos, and 3D content, and how such models may redefine the future of digital content creation.

The three PhD students in Changjian's group will present their leading projects of DancingBox, Sketch2Anim, and Sketch2Arti.