Dual Diffusion Bridges: Models & Mechanisms

• Dual Diffusion Bridges are diffusion-based transport constructions that model two endpoints to enable translation between paired source and target distributions.
• They utilize design patterns like shared-prior duality and bidirectional bridge duality, supporting applications in image translation, audio timbre transfer, and single-cell perturbation modeling.
• Advanced training objectives such as denoising bridge score matching and noise alignment optimize the deterministic and stochastic sampling processes for improved generation quality.

Dual diffusion bridges are diffusion-based transport constructions that organize generative modeling around two endpoints rather than a single data distribution and a fixed noise prior. In the current literature, the phrase is used in at least two distinct senses. One sense refers to two-endpoint bridge models that connect paired source and target distributions through a single bridge process, as in Denoising Diffusion Bridge Models (DDBMs) and the Bidirectional Diffusion Bridge Model (BDBM) (Zhou et al., 2023). A second sense refers to dual implicit bridges that concatenate two independently trained diffusion transports through a shared Gaussian prior, as in Dual Diffusion Implicit Bridges (DDIBs) and later latent-domain variants for audio and single-cell perturbation modeling (Su et al., 2022). Across both senses, the common objective is distribution translation under diffusion-style dynamics, with close ties to Doob’s $h$-transform, Schrödinger bridges, probability flow ODEs, and entropic or transport-based interpretations.

1. Terminological scope and historical placement

The literature does not use a single canonical definition of “dual diffusion bridges.” DDBM introduced a principled two-endpoint diffusion formulation for paired distributions but did not formalize “dual diffusion bridges” as a method name; its bridge is “dual” only in the sense of involving two marginals, endpoint conditioning, forward bridge dynamics, reverse bridge dynamics, and a probability flow ODE (Zhou et al., 2023). By contrast, BDBM explicitly formulates what the literature sometimes calls a dual diffusion bridge under the name Bidirectional Diffusion Bridge Model, where “dual” means one bridge, two directions, one model (Kieu et al., 12 Feb 2025). DDIB uses the term differently: it composes a source-domain diffusion model and a target-domain diffusion model through a common latent prior, so “dual” refers to two independently trained diffusion models and two bridge segments joined at the prior (Su et al., 2022).

A compact way to organize the field is to separate the main usages shown below.

Family	Core construction	Sense of “dual”
DDIB	source $\to$ Gaussian prior $\to$ target via two independently trained PF-ODEs	two models, two bridge segments
DDBM	one endpoint-conditioned bridge process between paired marginals	two endpoints, forward/reverse bridge dynamics
BDBM	one endpoint-conditioned bridge with analytic kernels in both time directions	one bridge, two directions, one network

This suggests that “dual diffusion bridges” is best treated as an umbrella notion for diffusion-based two-endpoint transport, rather than a single standardized architecture. A plausible implication is that the field has evolved from implicit dual composition through a shared prior toward explicit bidirectional modeling on one bridge process.

2. Mathematical foundations of two-endpoint bridge models

The basic bridge construction starts from a diffusion

$dx_t = f(x_t,t)\,dt + g(t)\,dw_t,$

then conditions the process to reach a terminal endpoint. In DDBM, conditioning on $x_T=y$ yields the bridge SDE

$$dx_t = f(x_t,t)\,dt + g(t)^2 \nabla_{x_t}\log p(x_T \mid x_t)\big|_{x_T=y}\,dt + g(t)\,dw_t, \qquad x_0 \sim p(x), \quad x_T = y.$$

The Doob correction

$$h(x,t,y,T) := \nabla_{x_t}\log p(x_T \mid x_t)\big|_{x_t=x,x_T=y}$$

steers the diffusion toward the endpoint. DDBM then defines a bridge process with endpoint pair distribution $p_{\mathrm{data}}(x_0,x_T)$ and derives the reverse conditional dynamics

$$dx_t = \Big[ f(x_t,t) - g^2(t)\big(s(x_t,t,y,T)-h(x_t,t,y,T)\big)\Big]dt + g(t)\,d\hat{w}_t, \qquad x_T=y,$$

together with the associated probability flow ODE

$$dx_t = \Big[ f(x_t,t)-g^2(t)\big(\tfrac12 s(x_t,t,y,T)-h(x_t,t,y,T)\big) \Big]dt.$$

Here

$\to$0

This makes the learned object an endpoint-conditioned bridge score, rather than the unconditional score field of ordinary diffusion (Zhou et al., 2023).

To obtain tractable supervision, DDBM chooses a Gaussian bridge marginal

$\to$1

with

$\to$2

Because this conditional is Gaussian,

$\to$3

and denoising bridge score matching recovers $\to$4. The formulation strictly generalizes ordinary diffusion: if $\to$5 is chosen as a Gaussian corruption of $\to$6, the bridge marginal collapses to the standard diffusion marginal, and the bridge reverse SDE and bridge ODE reduce to the usual reverse SDE and probability flow ODE (Zhou et al., 2023).

BDBM arrives at bidirectionality through a different route. It begins with endpoint-conditioned Gaussian marginals

$\to$7

with boundary conditions

$\to$8

Under this assumption, both

$\to$9

are Gaussian in closed form. The forward kernel is learned from a bridge Chapman–Kolmogorov equation, and the backward kernel follows by Bayes’ rule. The central consequence is that the same endpoint-conditioned process possesses analytic transitions in both time directions, which is the mathematical basis for BDBM’s bidirectional or dual character (Kieu et al., 12 Feb 2025).

3. Two dominant design patterns: shared-prior duality and bidirectional bridge duality

The shared-prior pattern is exemplified by DDIB. A source image $\to$0 is encoded to the common latent prior by the source model’s probability flow ODE,

$\to$1

and then decoded by the target model,

$\to$2

The resulting translation is deterministic in the continuous-time formulation and cycle-consistent up to discretization errors of the ODE solver. Theoretical interpretation is given as a concatenation of source-to-latent and latent-to-target Schrödinger bridges, that is, a form of entropy-regularized optimal transport through a shared reference prior (Su et al., 2022).

This shared-prior interpretation has been reused beyond images. In latent musical audio timbre transfer, dual diffusion bridges are implemented by training one latent diffusion model per instrument and composing source-to-prior and prior-to-target ODE solves. The bridge variable is a Gaussian latent prior,

$\to$3

and the terminal noise level $\to$4 becomes a practical control for the trade-off between melody preservation and timbre transfer (Mancusi et al., 2024). In unpaired single-cell perturbation estimation, DDIB is instantiated as dual conditional diffusion implicit bridges, where a control cell is inverted into a shared latent, then decoded under a perturbation condition to generate a counterfactual perturbed profile (Chi et al., 26 Jun 2025).

The second design pattern is explicit bidirectional modeling on a single bridge. BDBM uses one network

$\to$5

where $\to$6 denotes forward translation $\to$7 and $\to$8 denotes backward translation $\to$9. The paper’s intuition is that both directional predictors represent the same underlying Gaussian noise $dx_t = f(x_t,t)\,dt + g(t)\,dw_t,$0, so a single parameterization suffices (Kieu et al., 12 Feb 2025). This is a stricter and more explicit notion of duality than DDIB: one bridge, one shared network, two time directions.

A broader algorithmic version of this forward/backward pairing appears in Schrödinger-bridge solvers. “Aligned Diffusion Schrödinger Bridges” replaces standard forward/backward IPF-style estimation with paired endpoint supervision from aligned data, effectively collapsing the usual dual bridge estimation problem into a direct conditional bridge regression problem (Somnath et al., 2023). “Diffusion & Adversarial Schrödinger Bridges via Iterative Proportional Markovian Fitting” interprets practical alternation between forward-time and backward-time diffusion fitting as an Iterative Proportional Markovian Fitting procedure, combining reciprocal projection, Markovian fitting, and endpoint marginal correction (Kholkin et al., 2024). This suggests that duality in bridge modeling can be architectural, dynamical, or algorithmic.

4. Training objectives and parameterization strategies

Training objectives differ sharply across the major families. DDIB keeps the source and target models independent during training. Each domain-specific model is trained as a standard diffusion model on its own marginal, and duality appears only at inference through composition (Su et al., 2022). This modularity is precisely why DDIB scales to new domain pairs without joint retraining.

DDBM instead performs denoising bridge score matching. With training triples $dx_t = f(x_t,t)\,dt + g(t)\,dw_t,$1 and a bridge score network $dx_t = f(x_t,t)\,dt + g(t)\,dw_t,$2, the loss is

$dx_t = f(x_t,t)\,dt + g(t)\,dw_t,$3

and the minimizer satisfies

$dx_t = f(x_t,t)\,dt + g(t)\,dw_t,$4

In practice, DDBM adopts an EDM-style predict-$dx_t = f(x_t,t)\,dt + g(t)\,dw_t,$5 parameterization and generalized preconditioning coefficients $dx_t = f(x_t,t)\,dt + g(t)\,dw_t,$6 that depend on endpoint statistics $dx_t = f(x_t,t)\,dt + g(t)\,dw_t,$7. For unconditional generation it reuses EDM architectures; for pixel-space translation it uses ADM-style U-Nets, with endpoint conditioning by input concatenation (Zhou et al., 2023).

BDBM reduces its bridge KL objective to denoising regression on the shared latent Gaussian noise: $dx_t = f(x_t,t)\,dt + g(t)\,dw_t,$8 with $dx_t = f(x_t,t)\,dt + g(t)\,dw_t,$9. The paper also studies an endpoint-sum predictor $x_T=y$0, but reports that noise prediction works better, yielding lower FID and better diversity (Kieu et al., 12 Feb 2025).

A later refinement, Noise-Aligned Diffusion Bridge (NADB), diagnoses a specific failure mode in score-style bridge learning: as $x_T=y$1, the network input can become nearly deterministic while the regression target remains highly stochastic, producing endpoint underfitting. NADB replaces the I2SB-style bridge with

$x_T=y$2

where $x_T=y$3 is predicted by a separate mean network, and trains with

$x_T=y$4

The stated motivation is twofold: noise alignment resolves the magnitude mismatch near the target endpoint, while the mean network reduces the endpoint gap in Wasserstein-2 distance and corrects directional error (Gao et al., 27 May 2026).

A related optimization reformulation appears in “Simplified Diffusion Schrödinger Bridge,” which replaces original DSB targets by direct adjacent-state prediction,

$x_T=y$5

and then introduces terminal and flow reparameterizations analogous to $x_T=y$6-prediction and velocity prediction in ordinary diffusion and flow-matching systems (Tang et al., 2024).

5. Sampling, reversibility, and empirical behavior

Inference procedures reflect the underlying notion of duality. DDIB and its latent descendants use deterministic probability flow ODE transport. In the image formulation, exact cycle consistency holds in the zero-discretization limit; in practice, consistency is only approximate because ODE solvers introduce numerical error (Su et al., 2022). In the audio formulation, the same structure is retained with a Heun-based ODE solver, and the terminal noise level $x_T=y$7 directly modulates the content–style trade-off. For violin $x_T=y$8 flute transfer, for example, $x_T=y$9 gives DPD $$dx_t = f(x_t,t)\,dt + g(t)^2 \nabla_{x_t}\log p(x_T \mid x_t)\big|_{x_T=y}\,dt + g(t)\,dw_t, \qquad x_0 \sim p(x), \quad x_T = y.$$0 and FAD $$dx_t = f(x_t,t)\,dt + g(t)^2 \nabla_{x_t}\log p(x_T \mid x_t)\big|_{x_T=y}\,dt + g(t)\,dw_t, \qquad x_0 \sim p(x), \quad x_T = y.$$1, while $$dx_t = f(x_t,t)\,dt + g(t)^2 \nabla_{x_t}\log p(x_T \mid x_t)\big|_{x_T=y}\,dt + g(t)\,dw_t, \qquad x_0 \sim p(x), \quad x_T = y.$$2 gives DPD $$dx_t = f(x_t,t)\,dt + g(t)^2 \nabla_{x_t}\log p(x_T \mid x_t)\big|_{x_T=y}\,dt + g(t)\,dw_t, \qquad x_0 \sim p(x), \quad x_T = y.$$3 and FAD $$dx_t = f(x_t,t)\,dt + g(t)^2 \nabla_{x_t}\log p(x_T \mid x_t)\big|_{x_T=y}\,dt + g(t)\,dw_t, \qquad x_0 \sim p(x), \quad x_T = y.$$4; using $$dx_t = f(x_t,t)\,dt + g(t)^2 \nabla_{x_t}\log p(x_T \mid x_t)\big|_{x_T=y}\,dt + g(t)\,dw_t, \qquad x_0 \sim p(x), \quad x_T = y.$$5 recovers DPD $$dx_t = f(x_t,t)\,dt + g(t)^2 \nabla_{x_t}\log p(x_T \mid x_t)\big|_{x_T=y}\,dt + g(t)\,dw_t, \qquad x_0 \sim p(x), \quad x_T = y.$$6 and FAD $$dx_t = f(x_t,t)\,dt + g(t)^2 \nabla_{x_t}\log p(x_T \mid x_t)\big|_{x_T=y}\,dt + g(t)\,dw_t, \qquad x_0 \sim p(x), \quad x_T = y.$$7 (Mancusi et al., 2024).

DDBM offers both deterministic and stochastic sampling. Pure probability flow ODE sampling can yield overly averaged outputs because the endpoint is fixed and the trajectory follows an “expected” path. To mitigate this, DDBM introduces a hybrid sampler combining reverse-SDE Euler–Maruyama steps with Heun updates for the bridge ODE, and also generalizes the ODE with a guidance-like parameter $$dx_t = f(x_t,t)\,dt + g(t)^2 \nabla_{x_t}\log p(x_T \mid x_t)\big|_{x_T=y}\,dt + g(t)\,dw_t, \qquad x_0 \sim p(x), \quad x_T = y.$$8: $$dx_t = f(x_t,t)\,dt + g(t)^2 \nabla_{x_t}\log p(x_T \mid x_t)\big|_{x_T=y}\,dt + g(t)\,dw_t, \qquad x_0 \sim p(x), \quad x_T = y.$$9 Empirically, DDBM reports significant improvements over image-translation baselines and retains unconditional-generation quality when one endpoint is Gaussian noise. On CIFAR-10 it reports FID $$h(x,t,y,T) := \nabla_{x_t}\log p(x_T \mid x_t)\big|_{x_t=x,x_T=y}$$0 versus EDM $$h(x,t,y,T) := \nabla_{x_t}\log p(x_T \mid x_t)\big|_{x_t=x,x_T=y}$$1, and on FFHQ-64 it reports $$h(x,t,y,T) := \nabla_{x_t}\log p(x_T \mid x_t)\big|_{x_t=x,x_T=y}$$2 versus $$h(x,t,y,T) := \nabla_{x_t}\log p(x_T \mid x_t)\big|_{x_t=x,x_T=y}$$3 for EDM (Zhou et al., 2023).

BDBM samples in both directions from the same network. In forward translation $$h(x,t,y,T) := \nabla_{x_t}\log p(x_T \mid x_t)\big|_{x_t=x,x_T=y}$$4, it starts from $$h(x,t,y,T) := \nabla_{x_t}\log p(x_T \mid x_t)\big|_{x_t=x,x_T=y}$$5 and iterates toward $$h(x,t,y,T) := \nabla_{x_t}\log p(x_T \mid x_t)\big|_{x_t=x,x_T=y}$$6; in backward translation $$h(x,t,y,T) := \nabla_{x_t}\log p(x_T \mid x_t)\big|_{x_t=x,x_T=y}$$7, it starts from $$h(x,t,y,T) := \nabla_{x_t}\log p(x_T \mid x_t)\big|_{x_t=x,x_T=y}$$8 and iterates to $$h(x,t,y,T) := \nabla_{x_t}\log p(x_T \mid x_t)\big|_{x_t=x,x_T=y}$$9. Under the Brownian bridge specialization

$p_{\mathrm{data}}(x_0,x_T)$0

it reports strong bidirectional translation performance. On Edges $p_{\mathrm{data}}(x_0,x_T)$1 Shoes it reports FID $p_{\mathrm{data}}(x_0,x_T)$2, and on Edges $p_{\mathrm{data}}(x_0,x_T)$3 Handbags it reports FID $p_{\mathrm{data}}(x_0,x_T)$4, with the left/right numbers corresponding to reverse/forward directions (Kieu et al., 12 Feb 2025).

NADB is motivated by endpoint behavior rather than architectural duality, but its empirical results are important for bridge sampling quality. On ImageNet $p_{\mathrm{data}}(x_0,x_T)$5 deblurring with a uniform kernel and NFE $p_{\mathrm{data}}(x_0,x_T)$6, it reports improvement from I2SB FID $p_{\mathrm{data}}(x_0,x_T)$7 to NADB FID $p_{\mathrm{data}}(x_0,x_T)$8, with PSNR increasing from $p_{\mathrm{data}}(x_0,x_T)$9 to $$dx_t = \Big[ f(x_t,t) - g^2(t)\big(s(x_t,t,y,T)-h(x_t,t,y,T)\big)\Big]dt + g(t)\,d\hat{w}_t, \qquad x_T=y,$$0 and SSIM from $$dx_t = \Big[ f(x_t,t) - g^2(t)\big(s(x_t,t,y,T)-h(x_t,t,y,T)\big)\Big]dt + g(t)\,d\hat{w}_t, \qquad x_T=y,$$1 to $$dx_t = \Big[ f(x_t,t) - g^2(t)\big(s(x_t,t,y,T)-h(x_t,t,y,T)\big)\Big]dt + g(t)\,d\hat{w}_t, \qquad x_T=y,$$2 (Gao et al., 27 May 2026).

Point-cloud denoising provides a different sampling lesson. P2P-Bridge formulates a paired clean–noisy point-cloud bridge and shows that alignment is essential; without alignment the model fails to converge, and with alignment the deterministic OT-ODE-style path performs best in its ablation, suggesting that for denoising tasks with strong endpoint information, stochasticity may be less useful than in multimodal translation (Vogel et al., 2024).

6. Applications, relations to adjacent theories, and persistent limitations

Dual diffusion bridges now appear across several problem classes. Image-to-image translation is the central testbed for DDIB, DDBM, BDBM, NADB, and aligned Schrödinger-bridge methods (Su et al., 2022). Image restoration and deblurring are important for NADB, which interprets the bridge as connecting degraded and clean image distributions rather than noise and data (Gao et al., 27 May 2026). Musical audio timbre transfer adapts the shared-prior DDIB pattern to EnCodec latent space (Mancusi et al., 2024). Point-cloud denoising uses paired endpoint bridges with OT-like point alignment (Vogel et al., 2024). Unpaired single-cell perturbation estimation uses dual conditional DDIM inversion and decoding through a shared prior space (Chi et al., 26 Jun 2025). In function spaces, stochastic optimal control and Hilbert-space Doob transforms extend bridge ideas beyond finite-dimensional Euclidean state spaces, although that literature does not yet provide an explicit dual Schrödinger-potential system (Park et al., 2024).

Theoretical relations are similarly broad. DDBM explicitly links diffusion bridges to Doob’s $$dx_t = \Big[ f(x_t,t) - g^2(t)\big(s(x_t,t,y,T)-h(x_t,t,y,T)\big)\Big]dt + g(t)\,d\hat{w}_t, \qquad x_T=y,$$3-transform, Schrödinger bridges, and optimal transport or flow matching, and shows that in a small-noise VE limit its probability flow ODE reduces to the straight-line velocity $$dx_t = \Big[ f(x_t,t) - g^2(t)\big(s(x_t,t,y,T)-h(x_t,t,y,T)\big)\Big]dt + g(t)\,d\hat{w}_t, \qquad x_T=y,$$4, recovering OT-Flow-Matching and Rectified Flow (Zhou et al., 2023). Aligned Schrödinger-bridge work shows that once endpoint coupling information is available, forward/backward dual-IPF machinery can be replaced by direct conditional bridge regression (Somnath et al., 2023). IPMF shows that alternating forward-time and backward-time fitting is not merely a heuristic symmetry device but a hybrid of Iterative Proportional Fitting and Iterative Markovian Fitting (Kholkin et al., 2024).

Several misconceptions recur. One is that every two-endpoint bridge is automatically a bidirectional translator. DDBM does not train two explicit inverse maps; it learns a single bridge score conditioned on one endpoint, and its ODE is described as bidirectional only in the deterministic trajectory sense (Zhou et al., 2023). Another is that dual diffusion bridges necessarily solve the full Schrödinger bridge problem. DDIB uses a shared latent prior and deterministic inversion rather than forward/backward Schrödinger potentials, while DDBM learns a tractable bridge family rather than the exact entropic bridge (Su et al., 2022). A third is that bridge training inherits the endpoint conditioning of ordinary diffusion without pathology; NADB’s endpoint-underfitting diagnosis directly contradicts that assumption in restoration settings (Gao et al., 27 May 2026).

Limitations remain method-dependent but structurally consistent. Paired-data dependence is central in DDBM, BDBM, P2P-Bridge, and aligned SB formulations (Kieu et al., 12 Feb 2025). Shared-prior dual methods such as DDIB and Unlasting rely on the assumption that the source and target domains can be meaningfully coupled through the same latent prior (Su et al., 2022). BDBM’s strongest results are in discrete time, with continuous-time variants less competitive under limited training budgets (Kieu et al., 12 Feb 2025). Audio results are confined to monophonic single-instrument data (Mancusi et al., 2024). Function-space bridge theory still lacks the explicit dual potential machinery familiar from finite-dimensional Schrödinger bridges (Park et al., 2024).

Taken together, the literature supports a precise but plural view. “Dual diffusion bridges” denotes not one method but a family of diffusion-based two-endpoint transport constructions. Its principal variants are implicit dual composition through a shared prior, explicit bidirectional bridge learning on one endpoint-conditioned process, and forward/backward alternation in Schrödinger-bridge fitting. The unifying theme is that generation is recast as bridge transport between endpoint distributions, with duality expressed either through two models, two time directions, or two complementary projection steps.