In traditional 3D computer graphics, animation and deformation are deeply tied to a model's underlying discretization. If an animator creates a blendshape or a rigging system, the deformation instructions are typically stored as a list of vertex-specific offsets or weights. For example, an instruction might dictate that "vertex #4500 moves two units upward."
This vertex-bound paradigm creates a massive bottleneck: a deformation pipeline built for one 3D model cannot be easily transferred to another. If a target mesh has a different vertex count, a different triangulation pattern, or a non-humanoid structure, the animation instructions break entirely. Overcoming this limitation requires topology-agnostic mesh deformation, a paradigm shift that decouples motion and structural styling from specific mesh configurations.
Instead of treating a 3D model as an explicit collection of indexed vertices, topology-agnostic methods treat it either as a geometry embedded within a continuous spatial field or as a continuous manifold surface. By doing so, they allow complex animations, expressions, and geometric details to be seamlessly ported across diverse, unregistered, and stylized characters.
Volumetric and Flow-Based Warping
One of the most intuitive ways to bypass vertex dependencies is to deform the space containing the mesh rather than the mesh itself. If you warp the underlying coordinate space, any geometry embedded within that space will deform naturally along with it, regardless of its vertex density or triangulation.
[Source Motion / Video] ───> [Continuous Spatial Flow Field] ───> [Applied to Target Mesh Space]
│
▼
(Topology-Agnostic
Deformation)
This concept of spatial distortion fields was early leveraged by Xiaosong Yang et al. in Defoxel: A New Tool for Virtual Aesthetic Shaping. By treating deformation as a spatial force field grounded in solid mechanics, the authors decoupled the shaping logic from the underlying mesh. This allowed artists to define a specific "aesthetic signature" or tension within a volume of space and instantly apply it across entirely different geometries without manual reconfiguration.
When transferring dynamic performances, such as facial expressions from a video to a 3D avatar, this spatial mapping can take the form of dense coordinate flows. Supasorn Suwajanakorn et al. demonstrated this in What Makes Tom Hanks Look Like Tom Hanks by estimating 3D flow fields from driver videos and mapping those spatial displacements onto target meshes. Because the flow field is a continuous spatial mapping, it successfully decouples motion from identity, allowing a single source performance to control diverse 3D models.
In modern neural pipelines, this strategy has evolved to handle highly stylized or non-human avatars. For instance, Haoyu Wang et al. developed Versatile Face Animator: Driving Arbitrary 3D Facial Avatar in RGBD Space, which uses dense flow fields and surface-based weights to map 2D motion onto arbitrary 3D meshes. Because the framework avoids predefined vertex-based blendshapes, it achieves high-fidelity performance transfer across characters with completely varying vertex counts and structural configurations.
Landmark Interpolation via Radial Basis Functions
When dense spatial fields are unavailable, practitioners often rely on a sparse set of corresponding landmark points (e.g., the corners of the eyes or mouth). To propagate these sparse displacements smoothly across the rest of an arbitrary mesh, researchers use Radial Basis Functions (RBFs), such as thin-plate splines. RBFs act as global space interpolators: given a set of control point movements, they calculate a smooth, continuous deformation field for the entire surrounding space.
An early example of this is found in Human hand modeling from surface anatomy, where John P. Lewis et al. used RBFs to deform a generic 3D hand model to match an individual's unique hand shape based on 2D image features. Because the RBF interpolation operates on coordinate space rather than vertex indices, the underlying mesh resolution does not restrict the deformation logic.
This mesh-agnostic property makes RBFs highly effective for cross-platform applications, such as game engines. In MeInGame: Create a Game Character Face from a Single Portrait, Jiangke Lin et al. reconstructed facial shapes from single portraits and used RBF interpolation to map those shapes onto target meshes. This allowed the reconstructed faces to be deployed across different game engines with entirely different vertex structures.
The same mathematical intuition applies to generative modeling. Arman Maesumi in Explorable Mesh Deformation Subspaces from Unstructured 3D Generative Models interpreted the latent space transitions of 3D generators as continuous deformation fields. By using RBF interpolation, the author transferred semantic shape variations learned by an unstructured generator directly onto high-quality, detailed mesh structures, bypassing the need for a shared topology.
Abstracting Control Structures: Bones and Handles
Another powerful strategy for achieving topology independence is to condition deformations on an intermediary control structure—such as a skeleton, a set of bones, or abstract handles—rather than vertex-specific offsets. If a network learns to predict how a garment or skin deforms based solely on the orientation of a joint or handle, that model can generalize to any mesh bound to that control structure.
[Skeletal Joint Rotations / Handles] ───> [Learned Deformation Prior] ───> [Arbitrary Mesh Topology]
For body and clothing simulation, Yudi Li et al. introduced this concept in CTSN: Predicting cloth deformation for skeleton-based characters with a two-stream skinning network. By conditioning cloth deformation on skeletal joint rotations rather than a specific parametric body mesh (like SMPL), their architecture generalizes across diverse character topologies, including non-humanoid creatures and fantasy characters.
For facial modeling, Kai Yang et al. proposed ASM: Adaptive Skinning Model for High-Quality 3D Face Modeling. Instead of defining facial expressions through vertex-specific offsets, the authors defined deformations through bone-driven influences parameterized by a Gaussian Mixture Model. Because the deformation mechanism is decoupled from the vertex indices, the learned facial priors can be transferred seamlessly across different mesh resolutions and connectivity patterns without retraining.
To animate characters that lack a traditional skeleton entirely—such as a jellyfish or a ghost—Jiaxu Zhang et al. developed TapMo: Shape-aware Motion Generation of Skeleton-free Characters. They replaced rigid bones with adaptive mesh handles, creating a unified motion representation. This allowed highly diverse geometries to be animated within the same mathematical framework, completely removing the dependency on fixed human templates.
Continuous Surface Representations and Intrinsic Geometry
The most radical shift toward topology independence involves treating the 3D mesh not as a discrete graph of vertices and edges, but as a continuous geometric manifold. By leveraging intrinsic geometric operators—such as the Laplace-Beltrami operator or coordinate-free neural architectures like DiffusionNet—models can process shapes based on their continuous surface geometry rather than their specific triangulation.
In ScanTalk: 3D Talking Heads from Unregistered Scans, Federico Nocentini et al. utilized DiffusionNet to achieve speech-driven animation on raw, unregistered 3D scans. By treating the meshes as continuous geometric surfaces, their method eliminates the need for computationally expensive mesh registration, allowing direct animation of raw scans without losing fine surface details.
Similarly, prioritizing surface geometry over vertex indices enables robust auto-rigging pipelines. Dafei Qin et al. in Neural Face Rigging for Animating and Retargeting Facial Meshes in the Wild and Wenchao Ma et al. in RigAnyFace: Scaling Neural Facial Mesh Auto-Rigging with Unlabeled Data demonstrated that topology-agnostic architectures allow facial rigging models to generalize to "in-the-wild" data. These models can automatically generate high-quality blendshapes for stylized characters and monsters that do not conform to standard human templates, even when the meshes contain non-manifold or disconnected parts like separate eyeballs.
This continuous perspective also extends to transferring fine-grained surface details, such as wrinkles or textures. Marco Fumero in Nonlinear Spectral Geometry Processing via the TV Transform introduced a framework that decomposes geometry into scale-aware spectral bands using a Total Variation transform. Because this decomposition relies on intrinsic geometric properties rather than explicit vertex maps, it facilitates the transfer of complex surface details between meshes with entirely different topological connectivity.
Finally, this independence allows 3D animation to hook into powerful, class-agnostic 2D priors. In MotionDreamer: Exploring Semantic Video Diffusion Features for Zero-Shot 3D Mesh Animation, Petr Kellnhofer et al. optimized 3D mesh poses directly against semantic features from pre-trained Video Diffusion Models. Because the optimization targets continuous semantic features rather than a class-specific parametric model, it enables zero-shot animation across entirely diverse topologies—ranging from humans and animals to mechanical objects—within a single, unified framework.
Summary
By moving away from vertex-specific representations and embracing spatial fields, RBF interpolations, abstract control handles, and continuous surface processing, topology-agnostic deformation fundamentally changes how we interact with 3D shapes. It transforms animation from a rigid, asset-specific process into a fluid, portable property that can be applied universally across the geometric spectrum.
Key Foundational Work
Defoxel: A New Tool for Virtual Aesthetic Shaping (2009) — contributed 4 concept elements to this direction.
Human hand modeling from surface anatomy (2006) — contributed 2 concept elements to this direction.