Multisample Flow Matching: Straightening Flows with Minibatch Couplings

Continuous-time generative models, such as continuous normalizing flows and diffusion models, generate high-quality data by transforming random noise into realistic samples. However, standard methods pair noise and data independently during training. This independent sampling creates curved, tangled transformation paths, inflates gradient variance during optimization, and requires computationally expensive numerical simulations to generate individual samples.

The article introduces Multisample Flow Matching, a generalized simulation-free training framework designed to construct straighter, more efficient probability paths. The core objective is to evaluate whether coupling mini-batches of noise and data samples using optimal transport principles can accelerate training, reduce sample generation cost, and preserve the underlying data distributions.

The authors evaluated the framework across synthetic benchmarks and standard high-dimensional image datasets, including downsampled 32x32 and 64x64 ImageNet. The approach pairs batches of noise and data using techniques such as Batch Optimal Transport, Batch Entropic Optimal Transport, and faster ranking-based stable couplings. By formulating the training objective over joint distributions, the framework avoids costly adversarial min-max optimization and simulation during training while mathematically ensuring the target data distribution is preserved exactly.

The findings show substantial operational improvements over standard Flow Matching and diffusion models. In image generation benchmarks, the proposed batch couplings reduced the required sampling compute by 30% to 60% while achieving equivalent image quality, adding only a 0.8% to 4% computational overhead to training time. Batch-coupled models also exhibited significantly lower gradient variance, leading to faster training convergence. Furthermore, in high-dimensional synthetic transport problems with unknown cost functions, the method outperformed static mappings by correctly matching the true target distributions with low transport costs.

These performance gains allow machine learning deployments to significantly cut inference latency and computational expenses in production without degrading sample fidelity. Additionally, the approach provides a viable, computationally efficient tool for high-dimensional optimal transport problems in biological data modeling and computer vision. Because ranking-based stable couplings matched the performance of full optimal transport solvers at lower computational complexity, organizations can deploy this framework using lightweight pairing algorithms.

Organizations training continuous generative models should adopt mini-batch coupling strategies to reduce production inference costs and training times. Future research should evaluate scaling these couplings to full-resolution images, test their behavior across broader multi-modal applications, and explore optimal coupling methods when pairing across distributed multi-GPU environments.