GaussianMorphing: Mesh-Guided 3D Gaussians for Semantic-Aware Object Morphing

TL;DR

GaussianMorphing tackles semantic-aware 3D morphing from multi-view images by unifying mesh-guided deformation with 3D Gaussian Splatting. It binds Gaussians to a reconstructed mesh, learns semantic correspondences with a Graph Convolutional Network, and predicts a continuous morphing flow that updates Gaussian positions in tandem with mesh deformation. The approach optimizes a multi-objective loss balancing geometric fidelity, texture continuity, and semantic alignment, achieving coherent, textured 3D morphs without requiring labeled data. On the TexMorph benchmark, it outperforms state-of-the-art 2D and 3D methods, delivering improved structural stability, color fidelity, and edge continuity, with substantial perceptual gains indicated by user studies. This framework advances practical 3D morphing for visual effects and digital content creation by delivering high-quality textured morphs from ordinary multi-view inputs.

Abstract

We introduce GaussianMorphing, a novel framework for semantic-aware 3D shape and texture morphing from multi-view images. Previous approaches usually rely on point clouds or require pre-defined homeomorphic mappings for untextured data. Our method overcomes these limitations by leveraging mesh-guided 3D Gaussian Splatting (3DGS) for high-fidelity geometry and appearance modeling. The core of our framework is a unified deformation strategy that anchors 3DGaussians to reconstructed mesh patches, ensuring geometrically consistent transformations while preserving texture fidelity through topology-aware constraints. In parallel, our framework establishes unsupervised semantic correspondence by using the mesh topology as a geometric prior and maintains structural integrity via physically plausible point trajectories. This integrated approach preserves both local detail and global semantic coherence throughout the morphing process with out requiring labeled data. On our proposed TexMorph benchmark, GaussianMorphing substantially outperforms prior 2D/3D methods, reducing color consistency error ($ΔE$) by 22.2% and EI by 26.2%. Project page: https://baiyunshu.github.io/GAUSSIANMORPHING.github.io/

Figures (15)

• Figure 1: Our GaussianMorphing (left) takes input images of the source and target, reconstructs them into 3D Gaussian representations with surface meshes, and uses a mesh-guided strategy to generate intermediate shapes at timestamps $t\in[0,1]$. Unlike prior approaches, our method achieves Semantic-Aware Object Morphing with textured colors without relying on 3D input data. The comparison table (right) shows that our method uniquely generates fully textured 3D outputs directly from images, offering complete geometric and textural fidelity.
• Figure 1: Quantitative comparison of morphing methods evaluates structural similarity using the MSE of SSIM, color consistency with $\Delta E$, and edge continuity through EI.
• Figure 2: Method Overview. Our GaussianMorphing framework takes source $\mathcal{X}$ and target $\mathcal{Y}$ images as input. Surface meshes are extracted from 3D Gaussian Splatting (Sec. \ref{['sub:gaussion']}) and used with Gaussian points for geometry–texture alignment. Geometric features provide the correspondence matrix $\Pi_{XY}$ (Sec. \ref{['sub:Correspondence']}), and intermediate shapes are interpolated over time. Training relies on a joint loss (Sec. \ref{['sub:loss']}), yielding high-quality textured 3D morphing. (Up: Blender results; Down: correspondence visualization with Matplotlib.)
• Figure 3: Qualitative comparison of morphing methods on the benchmark dataset. Baselines include DiffMorpher zhang2024diffmorpher and FreeMorph cao2025freemorph for image morphing, NeuroMorph eisenberger2021neuromorph for texture-free 3D shape morphing, and MorphFlow tsai2022multiview for textured multi-view morphing without true geometry. Our method generates textured 3D morphing with geometric details directly from image inputs.
• Figure 4: Qualitative morphing results with non-isometric deformations, demonstrating robust interpolation under challenging geometric conditions (Up: synthetic datas; Middle: real-world scanned objects from GSO downs2022google; Bottom: real-world photos).