Denoising Diffusion Bridge Models

• DDBMs are generative models that construct stochastic diffusion bridges conditioned on both source and target distributions, enabling precise data transformations.
• They employ score-based learning and denoising objectives to estimate bridge scores, facilitating robust reverse-time sampling and accelerating convergence.
• Applications include image translation, medical imaging, speech enhancement, and time series forecasting, leveraging efficient samplers and deterministic or stochastic solvers.

Denoising Diffusion Bridge Models (DDBMs) are a class of generative and translation models that synthesize a stochastic (or deterministic) process to interpolate between two arbitrary endpoint distributions—rather than mapping noise to data as in standard diffusion models. By jointly pinning the initial and final states of a diffusion path, DDBMs enable principled modeling of coupled data distributions, with strong applications in image translation, inverse problems, medical imaging, speech enhancement, and more. The mathematical foundation primarily rests on the theory of diffusion bridges—stochastic processes constructed by conditioning a reference diffusion to connect two prescribed endpoint marginals, often cast in the Schrödinger bridge or Doob $h$-transform formalism.

1. Mathematical Foundations

A DDBM models a stochastic process $(x_t)_{t\in[0,T]}$ such that $x_0\sim p_0$ ("source" distribution) and $x_T\sim p_T$ ("target" distribution). Starting from a reference forward SDE,

$$d x_t = f(x_t, t)dt + g(t)d w_t, \quad x_0\sim p_0,$$

a diffusion bridge is constructed by conditioning the forward path to reach $x_T$ at $t=T$. By Girsanov's theorem (Doob $h$‐transform), the conditioned forward SDE becomes

$$d x_t = f(x_t, t)dt + g^2(t)\nabla_{x_t} \log p(x_T|x_t)dt + g(t)dw_t,$$

so that the bridge approaches $p_T$ at $(x_t)_{t\in[0,T]}$0 (Zhou et al., 2023). The reverse process (sampling from $(x_t)_{t\in[0,T]}$1 given $(x_t)_{t\in[0,T]}$2) yields a time-reversed SDE,

$(x_t)_{t\in[0,T]}$3

with $(x_t)_{t\in[0,T]}$4 the (unknown) bridge marginal and $(x_t)_{t\in[0,T]}$5 analytic for Gaussian references (Zheng et al., 2024).

For the special case of a Schrödinger bridge between an arbitrary target and an isotropic Gaussian (relevant in Bayesian computation), the process minimizes the path-space relative entropy to a reference diffusion, subject to endpoint constraints. The optimal bridge is induced by an SDE whose drift contains a time-dependent potential $(x_t)_{t\in[0,T]}$6, where $(x_t)_{t\in[0,T]}$7 solves a backward Kolmogorov equation with boundary data encoding the target and reference densities (Heng et al., 2023).

Explicit bridge marginals are analytically tractable for linear diffusions (e.g., with Ornstein-Uhlenbeck drifts), yielding closed-form Gaussian kernels:

$(x_t)_{t\in[0,T]}$8

where $(x_t)_{t\in[0,T]}$9, $x_0\sim p_0$0, $x_0\sim p_0$1 are schedule-dependent coefficients, typically functions of $x_0\sim p_0$2, $x_0\sim p_0$3, and signal-to-noise ratios (Zhou et al., 2023, Zheng et al., 2024, Aguila et al., 15 Oct 2025).

2. Score-Based Learning and Bridge Objectives

Central to DDBMs is estimating the "bridge score" function—i.e., the gradient $x_0\sim p_0$4—needed to drive the reverse process. This is achieved by denoising bridge score matching (DBSM):

$x_0\sim p_0$5

where the target score is available in closed form for the Gaussian bridge. The neural network $x_0\sim p_0$6 ingests $x_0\sim p_0$7, $x_0\sim p_0$8, and $x_0\sim p_0$9 (sometimes with context) and predicts the bridge score (Zhou et al., 2023, Aguila et al., 15 Oct 2025, Arslan et al., 2024).

Variants include direct regression in data-space (for deterministic bridges, e.g., in real-world image denoising (Yang et al., 2023)), or parameterizing the denoiser to directly output $x_T\sim p_T$0 and recovering the necessary score via the optimality conditions (the "pred-x" setting, (Zhou et al., 2023)). Training objectives can incorporate other losses, e.g., L1, L2, or adversarial penalties, particularly when aligning with application-specific metrics (Arslan et al., 2024).

Iterative Proportional Fitting (IPF), or alternating minimization in path-space KL, appears in the context of Schrödinger bridges and conditional simulation, where each IPF round alternates fitting forward and backward dynamics to match endpoint constraints (Heng et al., 2023, Shi et al., 2022).

3. Algorithms: Training and Sampling

DDBM training comprises sampling $x_T\sim p_T$1 pairs from the data (or synthesized endpoint distributions), drawing intermediate times $x_T\sim p_T$2, constructing $x_T\sim p_T$3, and minimizing the score-matching/regression objective.

A general high-level pseudocode, adapted from (Zhou et al., 2023, Aguila et al., 15 Oct 2025):

$x_T\sim p_T$5

Sampling (generation) in DDBMs is typically performed by numerically integrating the learned reverse-time SDE or the equivalent probability flow ODE. Modern advances include:

• Hybrid samplers (SDE plus ODE correction), with stochastic and deterministic steps (Zhou et al., 2023, Zheng et al., 2024).
• Implicit and accelerated samplers (DBIM, CDBM) that permit 4–50× faster sampling with similar fidelity, via discretized non-Markovian bridges or consistency training/distillation (Zheng et al., 2024, He et al., 2024).
• Deterministic "direct" sampling—especially when the bridge is linear—reducing to a single pass through the network (e.g., (Yang et al., 2023, Yang et al., 2024)).

4. Applications and Practical Instantiations

DDBMs have been adopted in numerous settings:

• Bayesian Sampling and Posterior Inference: Bridges between complicated posteriors (known up to normalization) and a normal reference, enabling sampling from otherwise intractable targets. The SB approach achieves convergence guarantees and reduced sampling steps compared to standard diffusion methods (Heng et al., 2023).
• Medical Image Translation and Counterfactual Synthesis: DDBMs have been shown to generate anatomically faithful healthy counterfactuals from pathological scans, outperforming DDPMs and supervised baselines in segmentation and anomaly detection (Aguila et al., 15 Oct 2025). Self-consistent recursive DDBMs, with a soft noise-prior schedule and recursive target estimation, yield high-fidelity modality translation in MRI and MRI-CT tasks (Arslan et al., 2024).
• Visual Navigation and Imitation Learning: By fusing priors (from heuristics, learned policies, or Gaussian samples) into the bridge, DDBMs dramatically accelerate inference for action sequence planning and increase success rates relative to standard diffusion-based and regression models (Ren et al., 14 Apr 2025).
• Time Series Forecasting: The S²DBM instantiates Brownian-bridge processes, with a lightweight prior map, for point-to-point and probabilistic time series forecasting. Performance exceeds or matches the SOTA on numerous benchmarks (Yang et al., 2024).
• Image/Point Cloud Denoising and Restoration: Deterministic and probabilistic bridges, with or without explicit stochasticity, have been applied to real-world image denoising (Yang et al., 2023) and 3D point cloud recovery from corrupted data via optimal transport-based couplings (Vogel et al., 2024).
• Speech Enhancement: Unified flow/diffusion bridge frameworks reveal equivalence between each bridge-step and predictive denoising, yielding competitive or superior SI-SNR and PESQ with efficient transformers and low computational footprint (Wang et al., 20 Feb 2026).

5. Design Space, Stochasticity, and Algorithmic Acceleration

Recent work has clarified the full design space of bridge-based models, showing that stochastic interpolant approaches, preconditioning, and endpoint mixing allow for explicit control over interpolation, variance, and sampling trade-offs (Zhang et al., 2024). Key observations:

• By tuning the bridge's transition kernel parameters $x_T\sim p_T$4, one can interpolate between deterministic ODEs, stochastic SDEs, and autoregressive or non-Markovian chains (Zheng et al., 2024, Yang et al., 2024).
• Adding base distribution noise enables output diversity modulation—quantifiable via Average Feature Distance (AFD)—critical for tasks like image-to-image translation (Zhang et al., 2024).
• Accelerated samplers, including DBIMs and consistency bridge models, reduce sampling costs by orders of magnitude by leveraging deterministic or consistency-trained solvers (Zheng et al., 2024, He et al., 2024).
• Specialized noise schedules (monotonic, non-zero endpoint variance) boost generalization and robustness to out-of-distribution endpoints (Arslan et al., 2024).

6. Theoretical and Practical Considerations

DDBMs inherit theoretical guarantees from the theory of Schrödinger bridges and entropic optimal transport: convergence under mild regularity, explicit path-space optimality, and tractable characterization for Gaussian-reference cases (Heng et al., 2023, Zhou et al., 2023).

Implementation-wise, U-Net or ResNet backbones with time-embedding and endpoint conditioning suffice for most vision tasks; efficient architectures (e.g., TF-GridNet, CSDI, PVCNN) enable scaling to speech or point cloud modalities (Wang et al., 20 Feb 2026, Yang et al., 2024, Vogel et al., 2024). Practical settings (number of steps, batch size, learning rate) match those of diffusion models, with modification as required for bridge-based training.

In high-dimensional applications, the choice of endpoint mapping, noise schedule, and loss weighting can substantially impact performance. Alternatives to direct score matching include data-space regression or adversarial augmentation, each suited to particular modalities or types of supervision (Arslan et al., 2024, Yang et al., 2023).

7. Extensions, Impact, and Open Challenges

DDBMs provide a unified lens on generative modeling between arbitrary data distributions, subsuming score-based models, flow-matching, and Schrödinger-bridge methods. They now underpin state-of-the-art solutions in image translation, denoising, inverse problems, visual imitation learning, point cloud restoration, and speech enhancement.

Despite advances in accelerated sampling and improved diversity/fidelity trade-offs, key challenges remain: (1) extending beyond Gaussian bridges to arbitrary endpoint distributions lacking analytic transition kernels; (2) developing robust, scalable architectures for diverse modalities; (3) incorporating richer conditioning (e.g., temporal, structural, or atlas priors); (4) analyzing predictive-equivalence limitations; and (5) scaling to real-time and high-resolution domains (Zhou et al., 2023, Arslan et al., 2024, Aguila et al., 15 Oct 2025, He et al., 2024, Zheng et al., 2024, Wang et al., 20 Feb 2026).

The framework’s versatility and strong empirical gains position DDBMs as a fundamental tool in machine learning for learning and controlling data-to-data transformations across scientific and engineering domains.