Extending diffusion-based 2D morphing’s smooth and semantically coherent transformations to 3D content

Establish techniques that extend the smooth and semantically coherent transformations achieved by diffusion-based 2D image morphing to 3D content, enabling plausible transformations of 3D object geometry and appearance.

Background

While deep diffusion models have dramatically improved the quality and diversity of 2D morphing—especially through attention and latent-space interpolation—directly carrying these benefits into 3D remains difficult due to challenges in modeling 3D geometry, texture evolution, and temporal consistency.

The paper positions MorphAny3D as a training-free framework addressing this gap by leveraging SLAT representations and attention-based fusion (MCA and TFSA) within Trellis to produce coherent and smooth 3D morphs across diverse categories.