MFGLab: A Unified Framework for Generative Modeling via Mean-Field Games
A new open-source PyTorch library, MFGLab, unifies twelve prominent generative models under the lens of mean-field games (MFGs). The library's primary API is a cost tuple that allows users to specify all twelve models—including Continuous Normalizing Flows, OT-Flow, Score-based Models, and Schrödinger Bridges—using four composable cost functions. The training loop, log-Jacobian computation, and reverse-ODE sampler are shared automatically. The paper also introduces DI-Flow, a novel cost design that employs a differentiable entropy functional to encourage mode coverage. Additionally, learning-based MFG solvers are provided that substantially outperform neural ODE-based methods. The work addresses two previously unexplored dimensions: the interaction term set to zero in many existing models, and the application of the rich family of MFG solvers to generative modeling. The paper is available on arXiv with ID 2607.23026.
Key facts
- MFGLab is an open-source PyTorch library for generative modeling.
- It unifies twelve prominent generative models under mean-field games.
- The API uses a cost tuple with four composable cost functions.
- Models include Continuous Normalizing Flows, OT-Flow, Score-based Models, and Schrödinger Bridges.
- DI-Flow is a novel cost design using differentiable entropy for mode coverage.
- Learning-based MFG solvers outperform neural ODE-based methods.
- The paper addresses the unexplored interaction term and MFG solvers in generative modeling.
- The paper is on arXiv with ID 2607.23026.
Entities
Institutions
- arXiv