Abstract
Standard diffusion-based image-to-image (I2I) translation models are inherently stochastic, leading to inconsistent results that can be problematic in high-fidelity applications like medical imaging or forensics. We present Dual-approx Bridge, a deterministic generative model that leverages the Brownian Bridge framework. By employing two specialized neural networks—a forward approximator and a reverse approximator—the model eliminates the need for random noise injection during sampling. Our approach achieves superior image quality and structural faithfulness with significantly fewer sampling steps compared to traditional GANs and stochastic diffusion models.
Introduction
In Image-to-Image (I2I) translation tasks, such as converting sketches to photos or performing super-resolution, the objective is typically to produce a specific, high-fidelity output corresponding to the input. Current methodologies present a significant trade-off. Generative Adversarial Networks (GANs) are deterministic but often suffer from artifacts and blurriness. Conversely, Diffusion and Bridge models produce high-quality imagery through iterative denoising but are inherently stochastic. This randomness means that identical inputs can yield varying outputs, a lack of "faithfulness" that is unacceptable in precision-critical fields.
While Probability Flow ODEs have been used to make diffusion models deterministic, they often result in a loss of fine detail when applied to Bridge models, which interpolate between two image domains rather than starting from pure noise. The Dual-approx Bridge addresses this by creating a predictable, high-quality path between domains.
Methodology
The Dual-approx Bridge builds upon the mathematics of Brownian Bridge diffusion. Unlike standard diffusion that transitions from an image to pure Gaussian noise, a Brownian Bridge directly models the diffusion path between two distinct domains (e.g., Domain A to Domain B).
The core innovation lies in the use of two distinct neural network approximators:
- Forward Approximator ($\epsilon_\theta$): Trained during the forward diffusive process to estimate the score function and the initial state of the image.
- Reverse Approximator ($Z_\phi$): Trained independently to approximate the scaled white noise used in the reverse (denoising) process.
By combining these networks, we derive a new iterative sampling formula that removes the requirement for injecting random noise at every step. This transforms the typically stochastic "random walk" of a bridge model into a single, optimal, deterministic route from the input image to the target output.
Experiments and Results
The model was evaluated on several standard computer vision benchmarks, including Cityscapes (label-to-photo), Edges2Handbags (sketch-to-photo), and super-resolution benchmarks (Urban100, BSD100).
Key Results and Consistency
Dual-approx Bridge demonstrated superior performance across multiple metrics. On the Cityscapes dataset, it achieved an FID of 48.70, significantly outperforming Pix2Pix (101.04) and CycleGAN (75.37). Furthermore, it improved the Structural Similarity Index (SSIM) by approximately 50% compared to previous state-of-the-art deterministic models.
A critical advantage of this model is its consistency. As shown in qualitative comparisons, while stochastic SDE-based models produce different colors or textures across multiple runs of the same input, the Dual-approx Bridge remains perfectly consistent.
Efficiency and Faithfulness
The authors identified a unique trade-off: while more sampling steps can improve perceived "realism" (FID), fewer steps (as few as 3 to 10) actually result in higher "faithfulness" to the ground truth (PSNR/SSIM). Because the model follows an optimal deterministic route, it can outperform complex diffusion models that require 1,000 steps, representing a massive leap in computational efficiency.
Conclusion
Dual-approx Bridge provides a robust solution for deterministic I2I translation, combining the high-quality generation of diffusion models with the predictability of GANs. Future work will focus on refining the initial sampling step to be 100% noise-free and exploring "Variance Exploding" (VE) formulations to further push the boundaries of image quality and detail restoration.
Summary updated