When we deform a 3D object—whether it is a digital character or a model of a human heart—we usually want to change its shape without breaking its structure. In geometry processing, this means preserving topology: a sphere should stay a sphere, and a donut should stay a donut.
Traditional neural networks often struggle with this. If you train a standard MLP to predict the new positions of points on a mesh, the network might accidentally push two points through each other, causing "self-intersections," or pull them apart so violently that the surface "tears."
A growing body of research, exemplified by Shanlin Sun, Daniele Baieri, and Qiang Ma, solves this by reframing deformation not as a single jump in position, but as a continuous flow governed by Neural Ordinary Differential Equations (Neural ODEs).
From Displacements to Velocity Fields
In a standard deformation model, you might calculate a displacement vector for every point: $P_{new} = P_{old} + \Delta P$. The problem is that $\Delta P$ is unconstrained. If $\Delta P$ is large, the mesh can easily collapse into a non-physical mess.
The "Neural Diffeomorphic Flow" approach, proposed in Topology-Preserving Shape Reconstruction and Registration via Neural Diffeomorphic Flow, changes the perspective. Instead of predicting the final destination, a neural network predicts a velocity field $v(x, t)$. To find the final position of a point, we integrate its path through this field over a "time" interval:
$$\frac{d\phi(x, t)}{dt} = v(\phi(x, t), t)$$
This integration ensures the transformation is a diffeomorphism—a mapping that is smooth, differentiable, and has a smooth inverse.
Why Flow Prevents Self-Intersection
The intuition for why this works is rooted in the uniqueness of solutions to ODEs. If the velocity field $v$ is sufficiently smooth, the paths (trajectories) of two different points can never cross.
Imagine a crowd of people moving through a hallway. If everyone is teleported to a new spot simultaneously (displacement), two people might end up in the same chair. But if everyone follows a continuous path and no two people can occupy the same space at the same time (flow), they can weave around each other, but they will never "overlap."
This property is critical for medical applications. In CortexODE, Qiang Ma and colleagues use this to reconstruct the highly folded surface of the brain's cortex. Because the flow is diffeomorphic, the folds of the brain can get very close to each other without the mesh ever self-intersecting, which is a requirement for valid clinical neuroimaging.
Preserving Topology and Volume
Because these flows are invertible, they are naturally "topology-aware." If you start with a simple sphere and warp it using a Neural ODE, the result is guaranteed to have the same connectivity as the sphere. Ka-Hei Hui and Ruihui Li utilize this in Neural Template to warp coarse templates into detailed meshes while ensuring that holes (like the space between chair legs) are preserved and never accidentally "healed" or merged.
Beyond topology, researchers are also concerned with volume preservation. When deforming a 3D model, it often looks "rubbery" or "thin" if the volume isn't maintained. Daniele Baieri and Filippo Maggioli address this in Implicit-ARAP by using invertible neural network architectures. These networks act as deformation functions that prevent common artifacts like thinning or bloating, even during extreme manipulations.
Bridging Fields and Meshes
A major trend in this area is the move toward Neural Fields—representing shapes as continuous functions (like Signed Distance Fields) rather than discrete triangles.
Methods like MIRNF apply these diffeomorphic flows to medical image registration, treating the deformation of an MRI scan as a continuous optimization problem. Similarly, Amir Vaxman and Changjian Li 0001 show in Canonical Consolidation Fields that by representing motion as a continuous velocity field, we can reconstruct smooth 4D "movies" from noisy, disconnected 3D point clouds.
By combining the mathematical rigor of classical geometry (diffeomorphisms) with the flexibility of neural networks (ODEs), this research direction provides a robust framework for modeling complex, non-rigid motion in everything from beating hearts to morphing digital shapes.