Flow Matching for Generative Modeling
By Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, Matt Le. 2022
5,473 citations recorded by the source corpus.
We introduce a new paradigm for generative modeling built on Continuous Normalizing Flows (CNFs), allowing us to train CNFs at unprecedented scale. Specifically, we present the notion of Flow Matching (FM), a simulation-free approach for training CNFs based on regressing vector fields of fixed conditional probability paths. Flow Matching is compatible with a general family of Gaussian probability paths for transforming between noise and data samples -- which subsumes existing diffusion paths as specific instances. Interestingly, we find that employing FM with diffusion paths results in a more robust and stable alternative for training diffusion models. Furthermore, Flow Matching opens the door to training CNFs with other, non-diffusion probability paths. An instance of particular interest is using Optimal Transport (OT) displacement interpolation to define the conditional probability paths. These paths are more efficient than diffusion paths, provide faster training and sampling, and result in better generalization. Training CNFs using Flow Matching on ImageNet leads to consistently better performance than alternative diffusion-based methods in terms of both likelihood and sample quality, and allows fast and reliable sample generation using off-the-shelf numerical ODE solvers.
Subfields: Computer Vision and Pattern Recognition, Statistical and Nonlinear Physics, Generative Models, Signal Processing
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Computer Vision and Pattern Recognition
- Luc Van Gool
- Dacheng Tao
- Ming-Hsuan Yang
- Trevor Darrell
- Xiaogang Wang
- Jitendra Malik
- Yi Yang
- Yu Qiao
In Statistical and Nonlinear Physics
- Sergey Levine
- Nicholas Matthew Boffi
- Chelsea Finn
- Michael S. Albergo
- Eric Vanden-Eijnden
- Eli Shechtman
- Yaron Lipman
- Brian Ichter
In Generative Models
- Andreas Blattmann
- Eli Shechtman
- Robin Rombach
- Axel Sauer
- Ziwei Liu
- Sergey Levine
- Tim Dockhorn
- Saining Xie
In Signal Processing
Citations in the corpus
- Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers (2023, positive): “Voicebox [58] and Audiobox [59] employed flow-matchingmethod[60]andachievedbetterspeechmodeling capability.”
- VALL-E 2: Neural Codec Language Models are Human Parity Zero-Shot Text to Speech Synthesizers (2024, methodology): “Voicebox [Le et al., 2024] and Audiobox [Vyas et al., 2023] used flow-matching method [Lipman et al., 2022] and achieved better speech modeling capability.”
- Audiobox: Unified Audio Generation with Natural Language Prompts (2023, methodology): “The authors of Lipman et al. (2023) presented an alternative called optimal transport (OT), which are conditional paths with constant directions and speeds.”
- Voicebox: Text-Guided Multilingual Universal Speech Generation at Scale (2023, methodology): “Lipman et al. [38] also presents another flow that recovers the path of diffusion models [57], which is more complex than the OT path.”
- Mean Flows for One-step Generative Modeling (2025, neutral): “Flow Matching [28, 1, 30] provides an intuitive and conceptually simple framework for constructing flow paths that transport one distribution to another.”
- π0: A Vision-Language-Action Flow Model for General Robot Control (2024, positive): “In contrast, our model employs a novel design that fine-tunes a VLM to produce actions via flow matching [32, 28], a variant of diffusion [20, 46].”
- Structured 3D Latents for Scalable and Versatile 3D Generation (2024, neutral): “ABO includes about 8K high-quality 3D models provided by Amazon.com.”
- π0.5: a Vision-Language-Action Model with Open-World Generalization (2025, positive): “Similarly to π 0 , we use flow-matching [50] to predict continuous actions in the final model.”
- SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics (2025, positive): “Differently from prior VLA architectures, we interleave cross and self-attention layers, thus using a conditional Flow Matching Transformer (Esser et al., 2024; Liu, 2022; Lipman et al., 2022) as v θ .”
- Transfusion: Predict the Next Token and Diffuse Images with One Multi-Modal Model (2024, methodology): “…discrete modalities [OpenAI et al., 2024, Dubey et al., 2024], diffusion models [Ho et al., 2020, Rombach et al., 2022a] and their generalizations [Lipman et al., 2022] are the state of the art for generating continuous modalities [Dai et al., 2023, Esser et al., 2024b, Bar-Tal et al., 2024].”
- Discrete Flow Matching (2024, methodology): “Surprisingly, Discrete FM exhibits similarities with the continuous Flow Matching (Lipman et al., 2022) approach proposed for continuous signals.”
- FLUX.1 Kontext: Flow Matching for In-Context Image Generation and Editing in Latent Space (2025, neutral): “Adversarial Diffusion Distillation Sampling of a flow matching model obtained by optimizing Equation (3) typically involves solving an ordinary or stochastic differential equation [30, 1], using 50–250 guided [16] network evaluations.”
- Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think (2024, positive): “(2) through existing ODE samplers ( e.g. , Euler sampler) starting from a random Gaussian noise ϵ ∼ N ( 0 , I ) (Lipman et al., 2022; Ma et al., 2024a).”
- Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success (2025, positive): “Notably, our enhanced fine-tuned OpenVLA policies outperform even the best fine-tuned π 0 policies [3, 36]—which benefit from a base model with larger-scale pretraining and a more sophisticated learning objective (flow matching [24])—as well as the Multi-modal Diffusion Transformer (MDT) [38].”
- InstaFlow: One Step is Enough for High-Quality Diffusion-Based Text-to-Image Generation (2023, methodology): “To address this challenge, we leverage Rectified Flow [1, 2], a recent advancement in generative models that utilizes probabilistic flows [17, 23, 24].”
- Bespoke Solvers for Generative Flow Models (2023, methodology): “When training a Diffusion or Flow Matching models, p t has the form p t ( x ) = p t ( x | x 1 ) q ( x 1 ) dx 1 , where p t ( x | x 1 ) = N ( x | α t x 1 , σ t I ) .”
- One Step Diffusion via Shortcut Models (2024, neutral): “Iterative denoising methods such as diffusion (Sohl-Dickstein et al., 2015; Ho et al., 2020; Song et al., 2020) and flow-matching (Lipman et al., 2022; Liu et al., 2022) have seen remarkable success in modelling diverse images (Rombach et al., 2022; Esser et al., 2024), video (Ho et al., 2022;…”
- F5-TTS: A Fairytaler that Fakes Fluent and Faithful Speech with Flow Matching (2024, methodology): “In particular, Flow Matching with Optimal Transport path (FM-OT) [32] is widely used in recent research fields not only text-to-speech [14, 33, 34, 35, 36] but also image generation [37] and music generation [38].”
- Understanding Diffusion Objectives as the ELBO with Simple Data Augmentation (2023, neutral): “This holds for, for example, the v - prediction loss of [Salimans and Ho, 2022] and flow matching with the optimal transport path [Lipman et al., 2022].”
- Hunyuan3D 2.0: Scaling Diffusion Models for High Resolution Textured 3D Assets Generation (2025, positive): “Then, we build up a dual-single stream transformer [45] on the latent space of our VAE with the flow-matching [53, 24] objective.”