TL;DR
This paper introduces Simulation-Free Entropic Semi-Unbalanced Optimal Transport (SF-ESUOT), a framework that scales dynamic entropic optimal transport to high-dimensional data without requiring diffusion pretraining. By pairing a stochastic optimal control formulation with the bridge "reciprocal property," the method achieves simulation-free training and single-step ($NFE=1$) generation, reaching competitive FID scores on CIFAR-10 and outperforming existing methods on image-to-image translation.
What Problem Does This Paper Address?
The Schrödinger Bridge (SB) problem—the dynamic equivalent of Entropic Optimal Transport (EOT)—finds a finite-time diffusion process that smoothly transitions a source distribution (like Gaussian noise) to a target distribution (like real images) while minimizing transport cost and entropy.
Despite their theoretical elegance, existing continuous SB methods suffer from two major bottlenecks:
- Severe computational burden: Existing formulations parameterize the drift vector field and rely on numerical simulation (solving SDEs/ODEs) during both training and inference, requiring tens to hundreds of steps ($NFE \ge 100$) and weeks of compute.
- Scalability and convergence issues: When mapping between distant, high-dimensional distributions (e.g., standard normal noise to complex images), standard SB models diverge unless initialized from a pretrained diffusion model. Furthermore, rigid marginal constraints make standard EOT sensitive to outliers and dataset imbalances.
Methods
The authors address these challenges by reformulating EOT as Entropic Semi-Unbalanced Optimal Transport (ESUOT) and deriving a simulation-free solver:
- Semi-Unbalanced Formulation: Instead of forcing exact marginal matching on the target, ESUOT relaxes the target constraint using an $f$-divergence penalty ($D_\Psi$). This adds stability and robustness to outliers while retaining EOT as a special case.
- Stochastic Optimal Control (SOC) & Dual Formulation: The authors derive the dynamical dual of ESUOT, proving that the inner control problem corresponds to a Hamilton-Jacobi-Bellman (HJB) PDE for a time-dependent value function $v_\phi(t, x)$.
- Decoupled Architecture with the Reciprocal Property: Instead of parameterizing a dynamic drift field, SF-ESUOT parameterizes:
- A static generator $T_\theta(x, z)$ mapping source samples directly to target predictions $\hat{y}$.
- A time-dependent value potential $v_\phi(t, x)$.
- Simulation-Free Training: By leveraging the reciprocal property of Schrödinger bridges, intermediate states $x_t$ can be sampled in closed form conditioned on endpoints $x$ and $\hat{y}$: $$x_t = (1-t)x + t\hat{y} + \sigma\sqrt{t(1-t)}\,\eta, \quad \eta \sim \mathcal{N}(0, I)$$ The potential $v_\phi$ is trained using a Physics-Informed Neural Network (PINN)-style residual loss on the HJB equation, while $T_\theta$ is updated against the HJB condition.
Datasets
- Synthetic Benchmarks:
- 2D toy tasks: Gaussian to 8-Gaussian ($G \to 8G$), Moon to Spiral ($M \to S$), and a 1D synthetic mixture with 2% outlier contamination.
- 50-dimensional Gaussian-to-Gaussian transport benchmark with known analytical solutions.
- Image Datasets:
- CIFAR-10 ($32 \times 32$): Evaluated for scratch generative modeling.
- CelebA ($64 \times 64$): Male $\to$ Female unpaired translation.
- AFHQ ($64 \times 64$): Wild $\to$ Cat unpaired translation.
- No new benchmarks were introduced.
Key Insights and Ideas
- Unbalancedness aids optimization: Formally, semi-unbalanced OT allows target distribution errors. Counterintuitively, relaxing the marginal constraint with a soft penalty (e.g., Softplus or KL divergence) actually produced better distribution matching (lower FID) on CIFAR-10 than strict EOT, acting as a regularizer during adversarial/dual training.
- Closed-form intermediate bridges bypass simulation: Traditional SB algorithms require costly SDE trajectory rollouts to acquire intermediate samples. Because optimal diffusion bridges condition into standard Gaussians given endpoints $(x, y)$, intermediate states can be generated instantaneously.
- One-step ($NFE=1$) evaluation: Because the generator $T_\theta$ directly predicts the target endpoint $y$, inference requires only a single forward pass without numerical ODE/SDE integration.
Key Results
- CIFAR-10 Generative Modeling:
- SF-ESUOT achieved FID 3.02 with $NFE = 1$ without pretraining, matching the state-of-the-art Schrödinger bridge model (IPF: FID 3.01, $NFE = 200$) that required pretraining.
- Standard scratch SB baselines (IPF, IMF, WLF) completely failed to converge ($\text{FID} \ge 100$).
- Reduced training wall-clock time from over 10 days down to approximately 4 days on 2 $\times$ RTX 3090Ti GPUs.
- Image-to-Image Translation:
- CelebA (Male $\to$ Female): Achieved an FID of 8.44, outperforming Neural Optimal Transport (11.96) and CycleGAN (12.94).
- AFHQ (Wild $\to$ Cat): Achieved an FID of 14.59, outperforming Diffusion Schrödinger Bridge Matching (DSBM $\ge 25$).
- Outlier Robustness: On 1D toy data with 2% contamination, SF-ESUOT successfully ignored out-of-distribution modes, whereas standard EOT overfitted to the outlier mass.
Limitations and Future Work
- Approximation precision vs. exact coupling: On the 50D Gaussian benchmark, SF-ESUOT showed slightly higher relative errors in mean, variance, and covariance ($\sim 3-4\%$) compared to simulation-heavy methods. The authors attribute this to the optimization difficulty of PINN-like residual objectives.
- Resolution constraints: Experiments were restricted to $32 \times 32$ and $64 \times 64$ resolutions; testing on higher-resolution datasets (e.g., CelebA-HQ $256 \times 256$) was omitted due to computational constraints.
- Future Directions: Developing improved optimization techniques for the HJB path constraint to enhance coupling fidelity without losing simulation-free efficiency.
Summary updated