Abstract:Deterministic flow models, such as rectified flows, offer a general framework for learning a deterministic transport map between two distributions, realized as the vector field for an ordinary differential equation (ODE). However, they are sensitive to model estimation and discretization errors and do not permit different samples conditioned on an intermediate state, limiting their application. We present a general method to turn the underlying ODE of such flow models into a family of stochastic differential equations (SDEs) that have the same marginal distributions. This method permits us to derive families of \emph{stochastic samplers}, for fixed (e.g., previously trained) \emph{deterministic} flow models, that continuously span the spectrum of deterministic and stochastic sampling, given access to the flow field and the score function. Our method provides additional degrees of freedom that help alleviate the issues with the deterministic samplers and empirically outperforms them. We empirically demonstrate advantages of our method on a toy Gaussian setup and on the large scale ImageNet generation task. Further, our family of stochastic samplers provide an additional knob for controlling the diversity of generation, which we qualitatively demonstrate in our experiments.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is: **Limitations and problems in the sampling process of deterministic flow models**.
Specifically, deterministic flow models (such as rectified flows) provide a general framework for learning deterministic transport mappings between two distributions, which are implemented as vector fields of ordinary differential equations (ODEs). However, these models have the following problems:
1. **Sensitivity to model estimation and discretization errors**: Due to the influence of model estimation and discretization errors, the performance of deterministic flow models is unstable.
2. **Inability to generate different samples based on intermediate states**: Deterministic flow models cannot generate different samples based on intermediate states, which limits their application range.
3. **Underestimation of the variance of the target distribution**: In some tasks, deterministic samplers will underestimate the variance of the target distribution, leading to bias.
To solve these problems, the author proposes a method to convert the basic ODE of the deterministic flow model into a family of stochastic differential equations (SDEs) with the same marginal distribution. This method allows a series of random samplers to be derived from a fixed deterministic flow model, thus achieving a continuous transition from deterministic to random sampling. By introducing additional degrees of freedom, this method can alleviate the problems of deterministic samplers and show better performance in experiments.
In addition, the author also shows how to improve the diversity of generated samples by controlling randomness and verifies the advantages of this method in the ImageNet generation task.
### Key contributions
1. **Proposing a family of flexible SDEs** (Theorem 1), these SDEs have the same marginal distribution as the given SDE or flow model, making it possible to explore different sampling schemes under a fixed model.
2. **Deriving new and existing special cases from Theorem 1** (Corollary 1.1 and Corollary 1.2), showing the wide applicability of Theorem 1.
3. **Studying a family of SDEs corresponding to the Gaussian flow model**, and showing flexible random sampling and controllable diversity generation in toy tasks and large - scale ImageNet generation tasks without retraining the model.
Through these contributions, this paper provides an effective method to improve the sampling process of deterministic flow models, thereby improving the robustness and generation quality of the models.