Technical Summary: Structured Diffusion Bridges: Inductive Bias for Denoising Diffusion Bridges

1. Problem Statement

Modality translation—the task of mapping information between heterogeneous domains (e.g., images to 3D, low-resolution to high-resolution)—is inherently under-constrained. While multiple cross-modal mappings can yield the same marginal distributions, existing Denoising Diffusion Bridge Models (DDBMs) and Latent Denoising Diffusion Bridge Models (LDDBMs) predominantly rely on fully paired datasets to define effective translations.

This exclusive reliance on paired supervision imposes a single data-driven constraint, implicitly enforcing semantic correspondence, distributional validity, and geometric consistency simultaneously. The authors argue that this strategy may be suboptimal, as the model might minimize reconstruction error without adhering to the underlying manifold structure. Furthermore, this approach fails in unpaired or semi-paired regimes where ground-truth correspondences are unavailable or scarce, limiting the applicability of diffusion bridges in real-world scenarios where dense pairing is often infeasible.

2. Methodology: Structured Diffusion Bridges (SDB)

The paper proposes Structured Diffusion Bridges (SDB), a framework that augments LDDBMs with explicit, composable structural constraints. Rather than treating paired supervision as a prerequisite, SDB treats it as an optional heuristic within a broader objective that characterizes the space of admissible solutions.

The core methodology involves learning a stochastic diffusion bridge in a shared latent space between source marginal $p_X$ and target marginal $p_Y$. To resolve the ambiguity of the joint distribution $p(x, y)$, SDB introduces three key components:

2.1. Marginal Matching (MM)

This constraint ensures the diffusion bridge terminates at the target marginal distribution $p_Y$.

• Mechanism: During training, target samples are encoded and corrupted. The bridge score function is trained via Denoising Score Matching (DSM).
• Heuristic: To reduce coupling ambiguity, a Winner-Takes-All (WTA) assignment is used. For each target sample, multiple conditioning candidates are drawn, and gradients are propagated only through the candidate that minimizes the loss. A capacity constraint prevents a few low-information samples from dominating the training.

2.2. Cycle Consistency (CC) as Bridge Reversibility

To enforce geometric consistency and invertibility without paired data, SDB introduces cycle consistency constraints at two levels:

• Endpoint Cycle Consistency: Enforces that the reconstructed latent representation $\hat{z}_0$ (obtained by translating $z_0 \to z_T \to \hat{z}_0$) is statistically consistent with the original $z_0$. This is formalized as minimizing $\| \hat{z}_0 - z_0 \|^2$.
• Trajectory-Level Cycle Consistency: Extends reversibility to intermediate diffusion states. It requires that the forward latent path $z_{t}^{X \to Y}$ and the reverse latent path $z_{T-t}^{Y \to X}$ remain consistent at matching diffusion times $t$. This regularizes the stochastic flow throughout the trajectory, preventing incoherent dynamics that might satisfy endpoint constraints but fail in the intermediate steps.

2.3. Unified Objective

The final training objective combines these heuristics with optional paired supervision:
$$\mathcal{L}_{total} = \mathcal{L}_{DSM} + \lambda_{end}\mathcal{L}_{end}^{cycle} + \lambda_{traj}\mathcal{L}_{traj}^{cycle} + \lambda_{pair}\mathbb{1}_{(x,y) \in D_{pair}}\mathcal{L}_{pair}$$
Where $\mathcal{L}_{pair}$ recovers the standard data-driven heuristic of DDBMs when paired data is available. The framework allows for seamless operation across unpaired ($\rho=0$), semi-paired ($0 < \rho < 1$), and fully paired ($\rho=1$) regimes.

3. Key Contributions

The authors summarize their contributions as follows:

1. Composite Objective Formulation: They formulate diffusion bridge training as a composite objective, treating paired supervision as one of several constraints rather than the sole driver.
2. Explicit Structural Regularization: They introduce marginal matching and multi-level cycle consistency (endpoint and trajectory), demonstrating that geometric enforcement improves translation quality even in fully paired settings.
3. Robustness in Low-Resource Settings: They demonstrate that SDB retains efficacy as paired data decreases, achieving near fully-paired quality with substantial relaxation in pairing requirements and remaining applicable in the fully unpaired regime.
4. Ablation Analysis: They provide a comprehensive analysis isolating the contributions of each constraint, highlighting their roles in shaping the diffusion trajectory and resolving coupling ambiguity.

4. Experimental Results

The method was validated on synthetic benchmarks and real-world tasks (Image Super-Resolution and Multi-view to 3D Voxel Generation) across varying levels of paired supervision ($\rho$).

• Synthetic Benchmarks: In controlled settings with known content-style structures, SDB showed that marginal matching alone yields poor content accuracy and high cycle error. Adding endpoint and trajectory cycle consistency significantly improved coupling quality and reversibility. SDB consistently outperformed paired-only baselines in semi-paired settings and maintained strong performance in unpaired settings.
• Image Super-Resolution (FFHQ $\to$ CelebA-HQ):
  • Fully Paired ($\rho=1$): SDB achieved a PSNR of 25.9, surpassing the LDDBM baseline (25.6).
  • Semi-Paired ($\rho=0.5$): SDB maintained competitive performance (PSNR 25.2) compared to fully supervised baselines.
  • Unpaired ($\rho=0$): SDB achieved a PSNR of 19.0, demonstrating meaningful translation capability where baselines like DiWa and LDDBM are inapplicable.
• Multi-view to 3D (ShapeNet):
  • SDB outperformed diffusion-based baselines (EDM, LDDBM) in paired and semi-paired regimes.
  • Crucially, while baselines failed or were inapplicable in the unpaired regime ($\rho=0$), SDB remained applicable, achieving an IoU of 0.549 and 1-NNA of 0.657, showing robustness to limited pairing.

5. Significance and Claims

The paper claims that diffusion bridges offer a flexible foundation for modality translation beyond fully paired data. By integrating principled geometric enforcement (via cycle consistency) with example-based methods, SDB addresses the under-constrained nature of modality translation.

The authors emphasize that:

• Dense pairing is not always necessary for semantic alignment; geometric constraints effectively capture the essential mechanics of translation.
• The framework allows for graceful degradation as pairing decreases, rather than catastrophic failure.
• Explicit structural constraints can enhance performance even when paired data is abundant, suggesting that current data-driven approaches may not fully exploit the problem's inherent structure.

The work positions SDB not as a replacement for paired methods, but as a more general framework that strengthens bridge-based translation via explicit structural constraints, making it viable for scenarios where paired data is scarce or non-existent.