Extended Flow Matching: a Method of Conditional Generation with Generalized Continuity Equation

Noboru Isobe,Masanori Koyama,Jinzhe Zhang,Kohei Hayashi,Kenji Fukumizu

2024-06-29

Abstract:The task of conditional generation is one of the most important applications of generative models, and numerous methods have been developed to date based on the celebrated flow-based models. However, many flow-based models in use today are not built to allow one to introduce an explicit inductive bias to how the conditional distribution to be generated changes with respect to conditions. This can result in unexpected behavior in the task of style transfer, for example. In this research, we introduce extended flow matching (EFM), a direct extension of flow matching that learns a ``matrix field'' corresponding to the continuous map from the space of conditions to the space of distributions. We show that we can introduce inductive bias to the conditional generation through the matrix field and demonstrate this fact with MMOT-EFM, a version of EFM that aims to minimize the Dirichlet energy or the sensitivity of the distribution with respect to conditions. We will present our theory along with experimental results that support the competitiveness of EFM in conditional generation.

Machine Learning,Analysis of PDEs,Functional Analysis,Optimization and Control,Probability

What problem does this paper attempt to address?

### What problem does this paper attempt to solve?

This paper aims to solve the problem that **generative models in conditional generation tasks lack explicit inductive bias**. Specifically, many existing flow - based generative models are unable to explicitly introduce inductive bias regarding conditional changes when dealing with continuous conditions. This may lead to unexpected behaviors in tasks such as style transfer.

#### Main problems:
1. **Limitations of existing flow models**: Most existing flow models are not designed to allow users to explicitly control how the conditional distribution changes with the condition. For example, in style transfer tasks, this lack of control may lead to generation results that are not as expected.

2. **The need to introduce inductive bias**: To improve this situation, researchers hope to introduce a method that enables the generative model to explicitly adjust the generation distribution according to the change of the condition, thereby avoiding unexpected results.

#### Solution:
To solve the above problems, the author proposes the **Extended Flow Matching (EFM)** method. EFM is a direct extension of the traditional Flow Matching (FM) method. By learning a matrix field, it can explicitly introduce inductive bias regarding conditional changes. Specifically:

- **The role of the matrix field**: EFM learns a matrix field \( u: \Xi \times D \to \mathbb{R}^{d \times (1 + k)} \), where \( \Xi \) is the product space of the conditional space and time, and \( D \) is the data space. This matrix field corresponds to a continuous mapping from the conditional space to the distribution space.

- **Introducing inductive bias**: By optimizing the objective function (such as minimizing Dirichlet energy), EFM can introduce inductive bias regarding conditional changes, ensuring that the generation distribution is less sensitive to conditional changes, thereby improving the controllability and stability of the generation results.

#### Experimental verification:
The author verifies the effectiveness of EFM through experiments, especially its performance in conditional generation and style transfer tasks. The experimental results show that EFM is competitive in these tasks and can better maintain the structural consistency of the generation results.

### Summary:
The main contribution of this paper is to propose a new generative model method - Extended Flow Matching (EFM), which can explicitly introduce inductive bias regarding conditional changes in conditional generation tasks, thereby improving the quality and controllability of the generation results.