When we look at two distinct objects—say, a sleek modern office chair and a classic wooden armchair—our minds can easily imagine a hybrid shape that smoothly blends the two. However, teaching computers to perform this morphing process has historically been a major challenge.
If you attempt to linearly interpolate between traditional discrete representations of these shapes, the geometry breaks down. In a voxel grid, parts of the first chair simply fade out while parts of the second chair fade in, creating a messy "ghosting" effect. In vector graphics or computer-aided design (CAD) models, interpolating the discrete sequences of drawing commands results in syntactic nonsense and broken files.
To solve this, modern geometric deep learning relies on a powerful concept: continuous latent manifold interpolation. By mapping complex, discrete geometric structures into a continuous, low-dimensional latent space, neural networks can learn a smooth manifold where every intermediate vector corresponds to a physically plausible, topologically consistent shape.
The Core Intuition: Why Discrete Interpolation Fails
To understand why latent interpolation works, we must first look at why explicit interpolation fails. Consider a 3D mesh or a pixel grid. These representations are discrete and localized. A vertex at coordinate $(x,y,z)$ represents a specific point in space, but it carries no semantic awareness of the broader object topology.
Early work by Alexey Dosovitskiy et al. in Learning to generate chairs with convolutional neural networks demonstrated that deep convolutional networks could invert this problem. By training on structured 3D data, the network learns an internal representation where intermediate points between known object categories represent physically plausible hybrid forms. Instead of morphing the pixels directly, we morph the recipe for the object within the network's internal space.
The Mechanism of Continuous Fields
The true breakthrough in achieving artifact-free shape morphing came with the shift from explicit representations (like voxels and meshes) to continuous implicit fields.
In Learning Implicit Fields for Generative Shape Modeling., Zhiqin Chen et al. introduced a framework where a shape is defined by a continuous decision boundary (an occupancy field). Similarly, [Jeong Joon Park](/author/jeong-joon-park) et al. developed DeepSDF: Learning Continuous Signed Distance Functions for Shape Representation, which represents surfaces as the zero-level set of a learned Signed Distance Function (SDF).
Discrete Voxel Interpolation:
Shape A (Solid) ----> Half-Faded A + Half-Faded B (Ghosting) ----> Shape B (Solid)
Continuous Implicit Field Interpolation:
Field A (z1) ----> Smoothly Warping Boundary Field (z_mid) ----> Field B (z2)
Why does this eliminate ghosting? When you interpolate between two latent vectors $z_1$ and $z_2$ in an implicit model, you are not blending spatial occupancies directly. Instead, you are interpolating the parameters of a continuous mathematical field. As the latent vector moves smoothly along a linear path, the network's learned function warps the underlying boundary continuously. The surface elements physically stretch, bend, and morph into the new geometry, preserving watertight properties and structural integrity throughout the entire transition.
Structuring the Latent Manifold
For interpolation to yield meaningful intermediate shapes, the latent space cannot just be a random collection of coordinates; it must form a well-behaved geometric manifold.
The authors of Shape Unicode: A Unified Shape Representation showed that by training a shared latent space across multiple modalities (such as voxels, point clouds, and multi-view images), the network captures a coherent semantic manifold. Proximity in this space directly correlates with geometric similarity.
To enforce this structure, researchers have explored different architectural constraints:
- Local Templates:
Kyle Genova et al.proposed Learning Shape Templates With Structured Implicit Functions, which decomposes shapes into a consistent set of local parameter elements. Averaging these local parameters guarantees that intermediate geometries remain structurally sound. - Probabilistic Continuity: In Diffusion Probabilistic Models for 3D Point Cloud Generation,
[Shitong Luo](/author/shitong-luo) et al.integrated a latent conditioning variable into a reverse diffusion process, ensuring that the underlying distribution remains smooth and continuous during point cloud generation.
Generalization Across Domains
The power of continuous latent manifolds extends far beyond standard 3D meshes. The same underlying principle—replacing brittle discrete tokens with continuous vector spaces—has unlocked advancements across diverse fields:
Vector Graphics and CAD Engineering
Traditional vector graphics (SVGs) and CAD models rely on sequential, discrete commands (e.g., "draw line to," "extrude circle"). Linearly interpolating these commands causes catastrophic syntax errors.
To bypass this, [Defu Cao](/author/defu-cao) et al. developed SVGformer: Representation Learning for Continuous Vector Graphics using Transformers, mapping mathematical paths into a continuous space to enable smooth font and icon morphing. Similarly, the authors of Revisiting CAD Model Generation by Learning Raster Sketch utilized a continuous latent space for CAD sketches, allowing designers to smoothly interpolate between complex geometric states, such as fluidly transitioning an aperture from a circle to a square.
Point Clouds and Radiance Fields
Point clouds are notoriously difficult to interpolate because they lack explicit connectivity. Ernst Röell et al. solved this in Generative Topology for Shape Synthesis and Point Cloud Synthesis Using Inner Product Transforms by mapping 3D structures into stable 2D directional projection images. Linear paths in this transformed latent space yield topologically consistent 3D shape morphs without chaotic particle scattering.
For novel view synthesis, [Yuelin Xin](/author/yuelin-xin) et al. demonstrated in Learning Unified Representation of 3D Gaussian Splatting that embedding 3D Gaussian primitives into a unified latent space ensures that interpolations and perturbations degrade gracefully, maintaining semantic coherence instead of causing rendering failures.
Medical Imaging and Physical Sciences
In medical applications, traversing these manifolds allows researchers to model biological variations. Amir H. Abdi developed AnatomyGen: Deep Anatomy Generation From Dense Representation With Applications in Mandible Synthesis, demonstrating that traversing the latent space between different patient landmark coordinates synthesizes anatomically realistic intermediate bone structures. For diagnostics, Rebeca Vétil et al. used latent space representations in Learning shape distributions from large databases of healthy organs: applications to zero-shot and few-shot abnormal pancreas detection to interpolate between healthy and diseased organ states, mapping out subtle clinical trends like progressive organ atrophy.
Even in chemistry, the concept holds true. Raffaella Buonsanti et al. showed in A holistic data-driven approach to synthesis predictions of colloidal nanocrystal shapes that mapping discrete nanocrystal geometries onto a continuous surface energy scale allows models to discover entirely new, physically viable crystal structures by targeting intermediate points on the energy continuum.
Summary
Whether morphing a chair, blending a human face (Zi-Hang Jiang et al.), or configuring neural network weights ([Olga Saukh](/author/olga-saukh) et al.), the core insight remains the same. By moving away from rigid, discrete descriptions and embracing continuous latent manifolds, we enable machine learning models to treat geometry not as a collection of isolated instances, but as a fluid, interconnected landscape.
Key Foundational Work
VConv-DAE: Deep Volumetric Shape Learning Without Object Labels (2016) — contributed 1 concept elements to this direction.
Learning to generate chairs with convolutional neural networks (2015) — contributed 1 concept elements to this direction.