ImageNet classification with deep convolutional neural networks
Alex Krizhevsky, Ilya Sutskever et al.
0
Citations
0
Influential Citations
—
Venue
2022
Year
… The goal of this work is to propose Flow Matching (FM), an efficient simulation-free approach to training CNF models, allowing the adoption of general probability paths to supervise …
Flow Matching addresses a critical bottleneck in training Continuous Normalizing Flows (CNFs): the reliance on expensive numerical ODE simulations during training. Traditional CNF training requires solving an ODE for each data point at every training step, which is computationally prohibitive for large-scale applications. This paper introduces a simulation-free objective that directly regresses the target vector field, making training significantly faster and more scalable.
The ability to use general probability paths is another major contribution. Previous methods like diffusion models are restricted to specific noise schedules, while Flow Matching allows the design of arbitrary paths that interpolate between noise and data. This flexibility opens the door to custom paths that may lead to better generative performance or faster sampling.
The paper demonstrates the effectiveness of Flow Matching on image generation tasks, particularly on CIFAR-10. The results show that Flow Matching achieves competitive Fréchet Inception Distance (FID) scores compared to existing CNF training methods, while requiring substantially less training time. For instance, the paper reports that Flow Matching can train a CNF in a fraction of the time needed for traditional CNF training, without sacrificing sample quality. The method also shows improved stability and ease of implementation.
Flow Matching has the potential to become a standard tool in generative modeling, offering a simpler and more efficient alternative to diffusion models and traditional CNFs. Its flexibility in designing probability paths could lead to new insights into the geometry of generative processes. Moreover, the simulation-free nature of the training makes it scalable to high-dimensional data, which is crucial for real-world applications such as image, audio, and video generation. The paper's clear theoretical foundation and strong empirical results make it a significant contribution to the field, likely inspiring further research on path-based generative models.
Alex Krizhevsky, Ilya Sutskever et al.
Ashish Vaswani, Noam Shazeer et al.
Douglas M. Bates, Martin Mächler et al.
Diederik P. Kingma, Jimmy Ba