machinestein/Y-Shaped-Generative-Flows

Y-Shaped Generative Flows

Most generative models learn V-shaped flows—each sample travels alone from prior to data along its own near-straight path. Effective, but blind to the fact that real data has hierarchical structure.

We introduce Y-Shaped Generative Flows: samples move together along shared "trunks" and only later branch into target-specific endpoints. Think rivers, trees, vascular systems, supply networks.

Key ideas:

• Models learn a shared representation path before specializing

• Better captures hierarchical relationships in data

• More efficient use of model capacity

• Natural fit for data with common structure + individual variation

Paper: https://arxiv.org/pdf/2510.11955 Code: https://github.com/machinestein/Y-Shaped-Generative-Flows