For years, generative modeling has been divided into two prominent continuous-time paradigms. On one side sit score-based diffusion models, which inject noise into data step-by-step and learn to reverse this degradation using Stochastic Differential Equations (SDEs). On the other side sit continuous-time normalizing flows and flow-matching models, which construct a deterministic velocity field to push a simple base distribution into a complex data distribution via Ordinary Differential Equations (ODEs).
While both approaches successfully generate high-quality samples, they have traditionally been treated as distinct mathematical frameworks with separate training objectives, sampling constraints, and trade-offs. A rapidly growing body of research unifies these paradigms under a single mathematical lens: stochastic interpolants and generalized diffusion bridges. This unification reveals that diffusion models and flow matching are not competing frameworks, but rather specific configurations along a continuous spectrum of transport mechanisms.
The Core Mechanism: Stochastic Interpolants
To understand why this unification is possible, we must look at how these models define the path between noise and data. Instead of defining a forward process implicitly by adding incremental noise over time, the stochastic interpolant framework constructs an explicit, time-dependent algebraic interpolation between two boundary distributions.
Let $x_0$ be a sample from a starting distribution (such as pure Gaussian noise) and $x_1$ be a sample from the target data distribution. A stochastic interpolant defines a trajectory $x_t$ across time $t \in [0, 1]$ by directly mixing $x_0$ and $x_1$ alongside a time-dependent noise profile. Because this path is explicitly parameterized, the velocity field required to move along the trajectory can be computed directly.
As demonstrated by Michael S. Albergo et al., this formulation exposes a profound mathematical duality. By training a neural network to predict this velocity field using a simple quadratic loss, we can construct a deterministic flow (an ODE). Crucially, the authors show that this exact same learned velocity field can be mathematically transformed into a score function for a diffusion process (an SDE).
This means that a single trained model can be sampled either deterministically or stochastically. A user can choose the deterministic ODE path for fast, few-step sampling, or switch to the stochastic SDE path to leverage the error-correcting properties of diffusion.
[ Stochastic Interpolant Path: x_t ]
|
+-------------------------+-------------------------+
| |
v v
[ Deterministic ODE ] [ Stochastic SDE ]
(Flow Matching / Velocity) (Diffusion / Score Function)
From Noise-to-Data to Arbitrary Bridges
Standard diffusion models are inherently anchored to a specific starting point: standard Gaussian noise. However, many real-world problems require transporting probability mass between two structured distributions—such as converting a blurry image into a sharp one, or translating a sketch into a photo.
By framing generative modeling through the lens of stochastic interpolants, we can generalize the endpoints to any two arbitrary distributions. This setup is known as a diffusion bridge.
Linqi Zhou et al. leverage this perspective to build Denoising Diffusion Bridge Models (DDBMs). By parameterizing the transport process as a generalized bridge, the authors show that a model can dynamically adapt its behavior between variance-preserving and optimal transport regimes based on the specific requirements of an image-to-image translation task.
This unified viewpoint reveals that many previously isolated domain-specific methods are actually variations of the same underlying mathematical structure. For instance:
- Conghan Yue et al. prove that Brownian bridges, variance-exploding paths, and variance-preserving paths can be unified under a single mean-reverting parametric lens for image restoration.
- Hyungjin Chung et al. demonstrate that various "direct" diffusion bridge methods differ primarily in their parameterization choices, allowing for a unified analysis of data consistency in inverse problems.
- Kaizhen Zhu et al. frame the entire diffusion bridge landscape as a generalized stochastic optimal control problem, providing a principled way to eliminate common artifacts like generation blurriness by tuning penalty and drift parameters.
Exploring the Continuous Design Space
When generative modeling is viewed as a continuous spectrum of interpolations, researchers can systematically break down the rigid constraints of older architectures to optimize training and sampling.
Decoupling Paths and Stochasticity
In traditional diffusion, the choice of the forward noise schedule strictly dictates both the training trajectory and the sampling noise. In contrast, the unified framework allows these components to be separated. Shaorong Zhang et al. (arXiv) and Shaorong Zhang et al. (NeurIPS) introduce Endpoint-Conditioned Stochastic Interpolants (ECSI). By decoupling the bridge path parameters from the sampling stochasticity, the authors unlock non-standard trajectories that intentionally alter the starting points, successfully resolving the long-standing problem of low output diversity in one-to-many image translation tasks.
Unifying Local and Global Objectives
The unification also spans the gap between step-by-step iterative generation and fast, single-step generation. Huijie Zhang et al. show that standard flow matching, shortcut models, and mean-flow objectives exist along a continuous spectrum. By adjusting a weighting parameter, a developer can balance local flow accuracy against global trajectory consistency, mitigating gradient conflicts during training.
Similarly, Jiayu Bai et al. unify standard diffusion models and consistency models under a single optimization objective. By tuning a single multiplier, the framework can transition smoothly from prioritizing local step-wise denoising to global single-step mapping, offering a principled way to trade off data efficiency and generation speed.
[ Local Step-wise Denoising ] <---- (Weighting Parameter) ----> [ Global Single-step Mapping ]
(Standard Diffusion/Flows) (Consistency Models)
Scaling Beyond Images
Because the stochastic interpolant framework operates on fundamental principles of probability transport rather than pixel-specific heuristics, its mathematical elegance allows it to scale across diverse domains and architectures:
- Infinite-Dimensional Spaces: Jianxin Zhang et al. extend rectified flows to infinite-dimensional Hilbert spaces. By utilizing flexible interpolation weights, their Functional Rectified Flow framework can generate continuous mathematical functions, such as complex fluid dynamics.
- Material Science: Philipp Höllmer et al. apply stochastic interpolants to crystal structure generation, allowing researchers to tune the specific transport paths required to map random distributions into stable inorganic materials.
- Reinforcement Learning: Yiyan Liang Edgar et al. leverage the framework to represent flexible time-dependent policies in offline reinforcement learning, significantly accelerating action sampling.
- Architectural Efficiency: The mathematical flexibility of the framework is highly compatible with modern non-Transformer backbones. For example, Vincent Tao Hu et al. combine a stochastic interpolant framework with a state-space model backbone (Zigzag Mamba), achieving high-quality image and video generation with superior memory scalability.
By shifting the paradigm from "reversing noise" to "interpolating distributions," the stochastic interpolant framework provides a unified, flexible, and mathematically rigorous foundation that clarifies why modern generative models work, while offering a systematic toolkit for building the next generation of architectures.
Key Foundational Work
Image Restoration Through Generalized Ornstein-Uhlenbeck Bridge (2024) — contributed 2 concept elements to this direction.
Exploring the Design Space of Diffusion Bridge Models via Stochasticity Control (2024) — contributed 2 concept elements to this direction.
ZigMa: A DiT-style Zigzag Mamba Diffusion Model (2024) — contributed 1 concept elements to this direction.