OmniMorpher: Zero-Inversion and Zero-Tuning 3D Morphing via Rectified-Flow Interpolation
3D morphological transition aims to generate smooth and semantically coherent interpolation sequences between distinct 3D assets, requiring the synchronous evolution of both geometric topology and surface textures. Existing methods, however, exhibit inherent limitations: point-cloud-based geometric mappings often suffer from structural distortions and texture mismatches, while approaches adapted from 2D image morphing rely on computationally cumbersome feature extraction pipelines. To overcome these challenges, we propose OmniMorpher, a versatile framework for high-fidelity 3D asset morphing. Grounded in the theoretical analysis of flow-driven models, we introduce a data-efficient, zero-inversion, and zerotuning flow interpolation strategy tailored for aligned 3D targets. OmniMorpher decomposes this complex task into two synergistic stages by leveraging a pre-trained 3D diffusion prior: (1) structural 3D latent encoding coupled with geometrically aligned transitions via optimized optimal transport, and (2) flow interpolation utilizing a spatially conditioned rectified flow model. Its flow-centric design is formulated to be portable to structured-latent rectified-flow priors that expose a compatible sparse encoder-decoder interface (a representation-compatibility claim, not an empirical cross-prior result). Extensive experiments, combining image-based, geometric, learned-judge, and human evaluations on externally sourced assets, show that OmniMorpher attains state-of-the-art geometric transition continuity and the highest learned-judge scores, while being significantly preferred by human participants, providing a practical route for flexible 3D content creation workflows.
Authors
- Shaoxu Li (ORCID: https://orcid.org/0000-0002-2686-1234)
- Ye Pan (ORCID: https://orcid.org/0000-0002-1447-6806)
- Yi Liu (ORCID: https://orcid.org/0000-0001-7575-2299)
Institutions
- Shanghai Jiao Tong University (CN)
- Baidu (China) (CN)
Publication Details
- Journal
- IEEE Transactions on Visualization and Computer Graphics
- Published
- 2026-11-20
- DOI
- https://doi.org/10.1109/tvcg.2026.3756111
- Primary Topic
- 3D Shape Modeling and Analysis
- Type
- article
- Field-Weighted Citation Impact
- 0.00