Morph Algorithms Overview

• Morph algorithms are a family of computational procedures that convert one representation to another while maintaining key domain-specific structures.
• They apply to diverse fields such as computer graphics for continuous deformations, graph drawing to maintain planarity and orthogonality, and biometric security via face morphing.
• Recent advances incorporate neural methods and adversarial optimization to enhance morphing quality, convergence speed, and adaptability across various applications.

Morph algorithms denote a heterogeneous family of computational procedures for transforming one representation into another while preserving domain-specific structure. In computer graphics and geometry, morphing is a continuous deformation between drawings, images, or surfaces; in graph drawing, it is constrained by invariants such as planarity, orthogonality, convexity, or geodesicity; in biometric security, a morph is a synthesized facial image containing biometric information of multiple identities; and in mathematical morphology, “morphological” algorithms implement dilation-, erosion-, and W-operator-based transformations rather than geometric interpolation (Goethem et al., 2018, Angelini et al., 2015, Bizzi et al., 10 Oct 2025, Blasingame et al., 2024).

1. Formal models and structural invariants

In graph drawing, a morph is a continuous transformation between two drawings of the same graph with the same combinatorial embedding. The literature distinguishes linear morphs, in which every vertex moves on a straight line at constant speed, from unidirectional morphs, in which all vertices move along parallel lines. For planar orthogonal drawings, the morph must preserve both planarity and orthogonality, and, ideally, linear complexity of the intermediate drawings (Goethem et al., 2018). For convex drawings of plane graphs, the requirement is stronger: every intermediate drawing must remain convex, or strictly convex when the inputs are strictly convex (Angelini et al., 2015).

In image and surface morphing, the canonical decomposition is warping plus blending. FLOWING formulates this explicitly for source and target media $f^0,f^1:\Omega\subset\mathbb{R}^n\to\mathbb{R}^m$, together with landmark correspondences, and identifies three desiderata for the deformation $\Phi$: uniqueness of trajectories, invertibility, and smoothness or energy minimization (Bizzi et al., 10 Oct 2025). In transformation optics, the source and destination are exact field images for two devices, and the intermediate result is an approximate wave picture obtained by morphing between them rather than recomputing a full finite-element solution (Aznavourian et al., 2014).

In biometric security, the term has a distinct meaning. A face morph is a single facial image synthesized from two different people so that a face recognition system may accept the same image as either person. Differential morph-attack detection then becomes a pairwise decision problem, typically comparing a trusted live capture against a suspect reference image (Chaudhary et al., 2021, Singh et al., 2019). A plausible implication is that “morph algorithm” is best treated as a family resemblance term rather than a single formalism.

2. Image, surface, and implicit-neural morphing

Neural image morphing has shifted from explicit correspondence engineering toward latent semantic interpolation. Neural Crossbreed replaces landmark specification and pixel warping with an encoder–decoder that disentangles content and style, learns semantic change from BigGAN-generated triplets, and supports separate control of shape and appearance transitions. The generator uses a content encoder $E_c$, a style encoder $E_s$, and a decoder $F$, with latent interpolations $c_\alpha=\mathrm{Lerp}(c_A,c_B,\alpha)$ and $s_\alpha=\mathrm{Lerp}(s_A,s_B,\alpha)$. Because supervision is obtained from interpolation in a pretrained BigGAN latent space, the method can handle large pose or camera differences without explicit correspondences (Park et al., 2020).

For 3D surfaces, conformal parameterization provides a different route to correspondence. “Conformal Surface Morphing with Applications on Facial Expressions” maps simply connected surfaces with a single boundary to the unit disk by Riemann conformal mapping, aligns them through an optimal Möbius transformation and a global thin-plate deformation, and then represents each surface by its mean curvature $H$ and conformal factor $\lambda$. Intermediate surfaces are generated by cubic spline homotopy of $(H,\lambda)$ on a unified mesh, followed by Gu–Yau surface reconstruction (Yueh et al., 2015). This replaces direct 3D coordinate interpolation with interpolation of intrinsic geometric data.

Topology-adaptive mesh deformation addresses the opposite regime, in which deformations may force self-intersections, merges, splits, or hole events. TransforMesh starts from a proper oriented mesh, detects self-intersections, identifies a seed triangle on the exterior using a winding-rule test, region-grows over exterior-valid faces, locally retriangulates intersecting triangles by constrained 2-D Delaunay triangulation, and stitches the resulting triangle soup back into a 2-D compact oriented manifold. The method is used for both surface morphing and 3-D reconstruction, with reported iteration times roughly from $\Phi$0 to $\Phi$1 per iteration on the tested machine (Zaharescu et al., 2020).

FLOWING recasts morphing as the construction of a differential vector flow rather than direct coordinate regression. Its flow-based parameterization is designed to encode continuity, invertibility, and temporal coherence directly into the architecture. The paper instantiates this with Neural ODEs and Neural Conjugate Flows, regularizes them with a thin-plate-like Jacobian penalty, and extends the framework to 2D images, generative latent blending, and 3D Gaussian Splatting morphing. It reports state-of-the-art morphing quality with faster convergence, while also noting a limitation: because invertibility is enforced by construction, the method cannot model topological changes or occlusions well (Bizzi et al., 10 Oct 2025). This contrasts sharply with TransforMesh, which is built precisely to accommodate topology change.

3. Morphing planar and toroidal graph drawings

For planar orthogonal drawings, the central problem is to preserve orthogonality and planarity while transforming one drawing into another. The algorithm of “Optimal Morphs of Planar Orthogonal Drawings” first unifies the drawings by subdividing edges so that both have the same number of segments, then replaces bends by vertices, introduces horizontal and vertical wires that encode topological order, and measures structural difference through the spirality of those wires. Zigzag-eliminating slides reduce the maximum absolute spirality by one in $\Phi$2 linear morphs, and rerouting plus cell-based cleanup keep intermediate complexity linear. The main theorem states that $\Phi$3 can be morphed to $\Phi$4 using $\Phi$5 linear morphs while maintaining $\Phi$6 complexity, resolving the open question left by Biedl et al. (Goethem et al., 2018).

For convex plane graphs, “Optimal Morphs of Convex Drawings” avoids contraction-based schemes and instead removes a suitable subgraph, recursively morphs the smaller convex graph, and reinserts the removed structure while maintaining convexity. The paper proves the existence of a convex unidirectional morph with $\Phi$7 morphing steps between any two convex drawings of the same $\Phi$8-vertex plane graph, and likewise a strictly-convex unidirectional morph with $\Phi$9 steps for strictly convex drawings. It also states that the linear bound is asymptotically optimal in the worst case (Angelini et al., 2015).

Barycentric interpolation provides a simpler paradigm for both planar and toroidal morphs. “Planar and Toroidal Morphs Made Easier” constructs planar morphs by interpolating the pair of weights associated with a single edge, entirely avoiding the classical edge-collapsing strategy. For two isomorphic convex planar drawings with the same convex outer face, it gives at most $E_c$0 unidirectional linear morphing steps; for arbitrary 3-connected planar straight-line drawings with the same convex outer face, it gives at most $E_c$1 such steps, both in $E_c$2 time. For the flat torus, it introduces a scaling strategy that transforms realizable barycentric weights into morphable weights, after which any intermediate torus drawing can be computed in $E_c$3 time (Erickson et al., 2021).

The toroidal case had previously required a more intricate inductive analysis. “How to Morph Graphs on the Torus” gives the first algorithm to morph isotopic essentially 3-connected geodesic embeddings on the Euclidean flat torus, producing $E_c$4 parallel linear morphing steps in $E_c$5 time. Its proof depends on a geometric analysis of 6-regular triangulations, the classification of bad hexagons into cats and dogs, a zipper base case, and a torus analogue of Tutte’s spring embedding theorem (Chambers et al., 2020). The later barycentric work may therefore be read as a simplification rather than a replacement.

4. Morphing as approximation in transformation optics

In transformation optics, morphing is used as a surrogate model for field visualization rather than as an exact solver. “Morphing for faster computations in transformation optics” computes exact source and destination field images, typically with COMSOL Multiphysics, places one-to-one corresponding control points on both images, and then uses morphing software such as Sqirlz Morph to generate intermediate field images. The paper describes two strategies: mesh warping with spline-based coordinate maps, and Delaunay triangulation with Voronoï-based localization of interior points (Aznavourian et al., 2014).

The physical setting is standard transformation optics. In the elliptical-cloak example, the map $E_c$6 leads to transformed tensors computed from the Jacobian $E_c$7 and

$E_c$8

The transformed medium becomes infinitely anisotropic at the inner boundary, so homogeneous Dirichlet or Neumann data are imposed there depending on polarization (Aznavourian et al., 2014).

The method is explicitly approximate. For cloaks, the error in $E_c$9 norm between the exact and approximate solutions is typically less than $E_s$0 percent if control points are judiciously chosen. The approach also works for rotators and concentrators, and it is used to suggest a hybrid device called a rotacon. It fails, however, for superscatterers generated by non-monotonic, space-folding transforms; in that case the $E_s$1 error is about $E_s$2 percent, and the paper reports differences “close to 0.5” in its normalized image metric. The computational advantage is substantial: in 2D, it takes less than $E_s$3 minute to deduce $E_s$4 images from the exact computations of the source and destination images, but the price is human intervention in control-point placement (Aznavourian et al., 2014).

5. Face morph generation, worst-case attacks, and differential detection

Morph generation for face recognition attacks has progressed from landmark blending to adversarial latent optimization and diffusion sampling. “Worst-Case Morphs using Wasserstein ALI and Improved MIPGAN” frames the attack in embedding space: if $E_s$5 is the face-recognition embedding and $E_s$6 a dissimilarity, the worst-case embedding $E_s$7 minimizes the maximum distance to the two source identities. The paper introduces Wasserstein ALI, a stabilized ALI variant with gradient penalties, and finetunes it with reconstruction and identity-aware losses so that generated morphs approach $E_s$8. It also transfers the same worst-case objective to improved MIPGAN, arguing that morphing attacks should be studied as adversarial optimization problems rather than only as image compositing procedures (Kelly et al., 2023).

Diffusion-based methods intensify that adversarial perspective. Greedy-DiM treats diffusion morphing as a stepwise optimization problem rather than a one-shot latent interpolation. Greedy-DiM-S searches over a discrete set of blend coefficients at every PF-ODE timestep, while Greedy-DiM* directly optimizes the predicted noise by gradient descent using an identity-based heuristic derived from ArcFace embeddings. On SYN-MAD 2022, Greedy-DiM* achieves $E_s$9 MMPMR on AdaFace, ArcFace, and ElasticFace, with $F$0, and also obtains $F$1 for $F$2 at $F$3. The paper nevertheless notes that it is slightly easier to detect than some other DiM variants in the MAD setting (Blasingame et al., 2024).

Differential morph detection addresses the converse task. “Differential Morph Face Detection using Discriminative Wavelet Sub-bands” applies an undecimated 2D DWT to obtain $F$4 mid- and high-frequency sub-bands, ranks them by Kullback-Liebler Divergence between bona fide and morph entropy distributions, selects the $F$5 most discriminative sub-bands, and trains a Siamese network on those bands. The reported best overall representation, BW-22, yields $F$6 EER on the universal test set and $F$7 EER on the unseen AMSL dataset, with dataset-specific D-EER values of $F$8 on MorGAN, $F$9 on LMA, and $c_\alpha=\mathrm{Lerp}(c_A,c_B,\alpha)$0 on VISAPP (Chaudhary et al., 2021).

In the Automated Border Control setting, a different differential architecture is used. “Robust Morph-Detection at Automated Border Control Gate using Deep Decomposed 3D Shape and Diffuse Reflectance” decomposes the live gate image and the eMRTD image with SfSNet into a diffuse reconstructed image and a quantized normal map, extracts an AlexNet fc7 feature of dimension $c_\alpha=\mathrm{Lerp}(c_A,c_B,\alpha)$1 from the diffuse branch, measures pairwise $c_\alpha=\mathrm{Lerp}(c_A,c_B,\alpha)$2 differences, trains linear SVMs, and fuses both feature scores and four camera scores by weighted sums. The fused system reports $c_\alpha=\mathrm{Lerp}(c_A,c_B,\alpha)$3 EER, compared with $c_\alpha=\mathrm{Lerp}(c_A,c_B,\alpha)$4 for the fused LBP-SVM baseline, with $c_\alpha=\mathrm{Lerp}(c_A,c_B,\alpha)$5 BPCER@APCER$c_\alpha=\mathrm{Lerp}(c_A,c_B,\alpha)$6 and $c_\alpha=\mathrm{Lerp}(c_A,c_B,\alpha)$7 BPCER@APCER$c_\alpha=\mathrm{Lerp}(c_A,c_B,\alpha)$8 (Singh et al., 2019).

6. Morphological algorithms and the broadened use of “MORPH”

A separate branch of the literature uses “morph” in the sense of mathematical morphology rather than geometric interpolation. “Deep Morphological Neural Networks” embeds differentiable approximations of dilation, erosion, opening, and closing into trainable neural layers, learns structuring elements by backpropagation, and introduces an adaptive morphological layer whose sign-controlled parameter selects dilation or erosion. On the reported classification benchmarks, the residual MNN achieves $c_\alpha=\mathrm{Lerp}(c_A,c_B,\alpha)$9 on MNIST, $s_\alpha=\mathrm{Lerp}(s_A,s_B,\alpha)$0 on SCGS, $s_\alpha=\mathrm{Lerp}(s_A,s_B,\alpha)$1 on GTSRB, and $s_\alpha=\mathrm{Lerp}(s_A,s_B,\alpha)$2 on the brain tumor dataset, improving to $s_\alpha=\mathrm{Lerp}(s_A,s_B,\alpha)$3 on GTSRB with dropout (Shen et al., 2019).

The discrete line of work goes further by preserving the Boolean-lattice structure exactly. “An Algorithm to Train Unrestricted Sequential Discrete Morphological Neural Networks” models each layer as a W-operator represented by a finite window and a Boolean characteristic function, and trains unrestricted sequential DMNN by a two-level stochastic lattice descent algorithm: one level over characteristic functions and one over connected windows. In the practical example of boundary recognition of noisy digits, the best learned two-layer USDMNN achieved minimum training error around $s_\alpha=\mathrm{Lerp}(s_A,s_B,\alpha)$4 and minimum validation error around $s_\alpha=\mathrm{Lerp}(s_A,s_B,\alpha)$5 (Marcondes et al., 2023).

In adjacent literatures, “MORPH” is frequently used as an acronym rather than a direct reference to morphing. MORPH for robot co-optimization introduces a differentiable hardware model proxy for joint optimization of design parameters and control policies in reinforcement learning, with CMA-ES used for design derivation (Bonato et al., 2023). MORPH for PDE modeling denotes a shape-agnostic autoregressive foundation model based on a Unified Physics Tensor Format, component-wise convolution, inter-field cross-attention, and axial attention, with transfer by either full fine-tuning or LoRA (Rautela et al., 25 Sep 2025). MORPH Wheel denotes a fully passive variable-radius wheel whose geometry and compliant structure implement torque-responsive transformation between direct-drive and radius-variation modes, with a range from $s_\alpha=\mathrm{Lerp}(s_A,s_B,\alpha)$6 to $s_\alpha=\mathrm{Lerp}(s_A,s_B,\alpha)$7 and threshold conditions based on force balance (Jang et al., 5 Feb 2026). This suggests a terminological broadening: “morph” now appears both as a label for continuous deformation algorithms and as a marker for adaptive transformation more generally.