Generative Molecular Morphing for Flexible-Size Design via Unbalanced Optimal Transport

Capturing the joint distribution of molecular properties and size is essential for steerable molecular generation. However, fixed-size diffusion- and flow-based models cannot adapt the number of atoms during generation, constraining their ability to represent this relationship. We introduce Morph, a flexible-size generative model for conditional and unconditional 3D molecular design. By dynamically adapting molecular size, Morph supports structural priors such as scaffolds and improves property steering. It matches state-of-the-art fixed-size models in de novo generation while better recovering the training distribution, and generates valid molecules in out-of-distribution size and property regimes. Thus, Morph combines competitive sample quality with flexible molecular generation.