MorphAny3D: Advances in 3D Morphing

• MorphAny3D is a comprehensive framework for transforming 3D models into semantically coherent and geometrically plausible intermediates.
• It leverages structured latent attention, transformer-based flows, and neural as well as statistical methods to blend, register, and edit complex meshes.
• The system achieves state-of-the-art performance in morph quality, topology adaptation, and real-time local editing for diverse 3D applications.

MorphAny3D encompasses recent algorithmic advances and frameworks for 3D morphing: the task of transforming one three-dimensional object into another via a sequence of semantically coherent, geometrically plausible intermediates. This includes category-crossing morphs, fine-scale editing, correspondence, topology preservation or adaptation, and style/appearance blending. MorphAny3D systems span mesh-based, diffusion-based, statistical, and neural approaches, leveraging innovations in structured latent representations, attention-modulated generative models, geometry-aware mesh processing, and locally controllable autoencoders. Key developments include mechanisms for high-quality 3D sequence generation, handling of complex topology evolution, real-time localized editing, and robust mesh correspondence.

1. Structured Latent Attention for Semantic 3D Morphing

MorphAny3D, as formalized in (Sun et al., 1 Jan 2026), leverages Structured Latent (SLAT) representations—sparse sets of paired latent codes and voxel-aligned positions, $(z_i, p_i)$—as the foundational data structure for 3D morphing. These are generated via a two-stage transformer-based flow: a Sparse Structure (SS) predictor yields surface voxel locations, followed by a SLAT flow that assigns local geometry-appearance codes to each position. This explicit, lattice-regular SLAT representation enables direct, correspondence-aware fusion and blending of source and target shapes within neural 3D generative pipelines.

Within the generative process, Morphing Cross-Attention (MCA) and Temporal-Fused Self-Attention (TFSA) mechanisms are injected into the transformer blocks. MCA blends the respective cross-attention responses of source and target features via an explicit, time-varying deformation schedule:

$$f^n_{MCA} = (1-\alpha^n)\,\text{Attn}(Q^n, K^s, V^s) + \alpha^n\,\text{Attn}(Q^n, K^t, V^t)$$

where $\alpha^n$ interpolates the morphing timeline and $Attn$ denotes the standard multi-head attention. This approach avoids artifacts from blending latent tokens naively and maintains semantic patch correspondence throughout morph sequences.

TFSA augments per-frame consistency by blending current-frame self-attention outputs with those from the previous frame:

$$f^n_{TFSA} = (1-\beta)\,\text{Attn}(Q^n, K^n, V^n) + \beta\,\text{Attn}(Q^n, K^{n-1}, V^{n-1})$$

where $\beta$ is fixed (e.g., 0.2), promoting temporal smoothness without sacrificing detail or introducing temporal drift.

An additional orientation correction procedure, applied after the SS stage, mitigates pose ambiguity by selecting the minimal Chamfer distance alignment among canonical yaw rotations, enforcing global structural correspondence and minimizing sudden pose flips commonly observed near the center of the morph sequence.

2. Statistical and Machine-Learned Mesh Morphing Pipelines

Geometric morphing and mesh correspondence are addressed by mesh-based frameworks, most notably in the context of finite element models and medical/engineering domains. The pipeline instantiated in (Andreassen et al., 2022) utilizes Generalized Regression Neural Networks (GRNN) as a powerful alternative to classic radial basis function (RBF) schemes.

In MorphAny3D's GRNN mesh-morphing variant, the pipeline proceeds through:

• Affine pre-alignment using Iterative Closest Point (ICP).
• Multi-scale mesh reduction for computational tractability.
• Computation of nodewise displacement vectors via nearest-neighbor search across source and target reduced clouds.
• Training a GRNN with Gaussian RBF kernels to interpolate the displacement field:

$$f(x) = \frac{\sum_i y_i \phi(\|x-x_i\|)}{\sum_i \phi(\|x-x_i\|)},\quad \phi(r) = \exp\left(-\frac{r^2}{2\sigma^2}\right)$$

• Iterative application to both reduced and full-resolution meshes until convergence in displacement magnitude.
• Optional overclosure correction to address mesh interpenetrations, with user-controlled adjustment weights and minimum gap parameters.

These methods achieve high morphing accuracy and mesh quality (e.g., median Hausdorff error as low as 0.28 mm for femur models), outperforming both classical RBF and Coherent Point Drift (CPD) approaches in convergence, runtime, and geometric fidelity.

3. Topology-Adaptive Mesh Morphing

Explicitly mesh-based MorphAny3D systems can achieve topology-adaptive shape evolution using frameworks such as TransforMesh (Zaharescu et al., 2020), capable of handling self-intersection removal, mergers, splits, hole formation, and dynamic adaptation of surface genus.

The core algorithm combines:

• Intersection detection via Axis-Aligned Bounding Box (AABB) trees for sub-quadratic collision identification.
• Valid “outside” face detection using generalized winding number tests.
• Local constrained Delaunay triangulation and region-growing for partial/interior face re-triangulation and recombination.
• Geometric and valence-based remeshing (edge splits, collapses, swaps) for mesh quality preservation.
• Laplace–Beltrami or higher-order smoothing for mesh relaxation after topology-altering events.

Dynamic topology modifications—such as splitting a torus or merging disconnected shells—are fully automated, with the explicit mesh always restored to a 2-manifold after each morphing step. This enables precise, large-scale deformations and morphs not easily achievable with implicit or volumetric methods while retaining adaptive resolution and texture fidelity.

4. Morphable Models: Single-Scan to Locally Adaptive Approaches

Statistical morphable models for 3D shape categories have evolved from multi-scan principal component analysis to flexible, kernel-based and locally adaptive approaches.

Gaussian Process Morphable Models (GPMMs) (Sutherland et al., 2020) allow the construction of 3D morphable models from a single scan by treating shape deformation fields as draws from GP priors with RBF kernels:

$$g_s \sim GP(\mu_s, K_s),\quad K_s(x,y) = I_3\cdot\Sigma_0(x,y),\quad \Sigma_0(x,y)=a_s\Sigma_{s,A_s}(x,y)+b_s\Sigma_{s,B_s}(x,y)+c_s\Sigma_{s,C_s}(x,y)$$

Such models support straightforward synthesis, registration, and inverse graphics tasks, and multi-scan nonparametric mixtures without requiring dense correspondence or category-specific tuning.

Locally Adaptive Morphable Models (LAMM) (Tarasiou et al., 2024) further extend this concept by introducing region-token-based autoencoders capable of direct, real-time local geometry editing. LAMM exposes user-controlled sparse vertex displacements via learned MLPs for each region, seamlessly integrating local edits with the globally encoded mesh structure in a highly efficient, single-forward computational graph.

5. Implementation Strategies, Performance, and Evaluation

MorphAny3D frameworks provide both efficient algorithms and reference implementations. Matlab-based pipelines (Andreassen et al., 2022) feature modular routines for mesh alignment, reduction, GRNN training, and overclosure adjustment. Transformer-based diffusion pipelines (Sun et al., 1 Jan 2026) integrate MCA and TFSA during inference on pretrained image-to-3D models without further training, supporting flexible applications including style transfer, dual-target blending, and disentangled morphing.

Quantitative experiments demonstrate state-of-the-art morphing quality:

• FID scores as low as 111.95 (best-in-class), with high temporal consistency (PDV = 0.0006) and strong user preference (UP = 86.73%) for cross-category morphing (Sun et al., 1 Jan 2026).
• Real-time mesh editing performance, with 12k-vertex meshes processed at $>$60 fps on CPUs using LAMM, surpassing previous GCN or SpiralNet++ methods by an order of magnitude (Tarasiou et al., 2024).
• Robust convergence and mesh quality over large deformations and topological events with TransforMesh, handling objects up to 50k faces in seconds (Zaharescu et al., 2020).

Best practices include careful affine pre-alignment, multi-scale processing, validation via Hausdorff error and element distortion, and, in the mesh domain, regular remeshing and constraint enforcement.

6. Limitations, Advanced Applications, and Generalization

MorphAny3D frameworks exhibit several limitations:

• Structured Latent-based approaches may inherit generator-specific artifacts (e.g., fine-geometry noise), and orientation correction, while effective for large pose jumps, may leave minor pose inconsistencies (Sun et al., 1 Jan 2026).
• Mesh-based approaches face storage/computation bottlenecks, especially in high-curvature regions (mitigated by adaptive reduction and sparse kernels), and potential challenges in extreme genus changes (Andreassen et al., 2022, Zaharescu et al., 2020).
• Real-time local editing is generally limited by the granularity of control-vertex selection and region tokenization in LAMM (Tarasiou et al., 2024).

Extensions include style transfer by swapping target SLATs for texture interpolation, dual-target morphing (separately blending structure and appearance), multi-body assembly processing, and integration with 3DGS/NeRF-style scene representations. The core attention and latent-blending strategies of MorphAny3D generalize to any transformer-based 3D generator exhibiting a two-stage structure-detail split with spatially indexed latent codes; successful demonstrations include Hi3DGen and Trellis text-to-3D pipelines (Sun et al., 1 Jan 2026).

7. Concluding Perspective and Outlook

MorphAny3D unifies algorithmic, neural, and statistical paradigms for 3D morphing, providing robust, high-quality, and application-agnostic solutions for shape interpolation, mesh editing, and topological evolution. Its advances in structured attention-based fusions, geometry- and correspondence-aware networks, topology-adaptive mesh processing, and fast, locally controllable editing constitute the foundation for next-generation 3D modeling and animation pipelines. The availability of open-source implementations and empirical validation across a range of morphing and registration tasks position MorphAny3D frameworks as central tools in scientific, engineering, and creative 3D workflows (Zaharescu et al., 2020, Andreassen et al., 2022, Sun et al., 1 Jan 2026, Sutherland et al., 2020, Tarasiou et al., 2024).