From Slow Diffusion to Fast Inference: Advances in Diffusion Bridge Models through Inverse Bridge Matching Distillation

Diffusion Bridge Models (DBMs) have emerged as a promising extension of diffusion models in the field of image-to-image transformation. They enable impressive results in tasks such as super-resolution, image restoration, and the generation of images from sketches. A significant drawback of DBMs, similar to many modern diffusion and flow models, however, is the slow inference speed. This delay considerably limits the practical applicability of the models, especially in real-time applications.

A new approach to solving this problem is "Inverse Bridge Matching Distillation." This innovative distillation technique aims to transfer the knowledge of a complex and slow DBM teacher model to a significantly faster student model. The core of this method lies in the inverse bridge matching formulation, which allows for defining a practicable optimization process for the distillation.

A crucial advantage of this technique over previous DBM distillation methods is its versatility. It can be applied to both conditional and unconditional DBMs. Conditional DBMs use additional information, such as a sketch, to control image generation, while unconditional DBMs generate images without additional input. Inverse Bridge Matching Distillation allows both types of models to be distilled effectively.

Furthermore, the method is characterized by the ability to distill models in a single step, using only corrupted images for training. This significantly simplifies the distillation process and reduces the computational cost. In contrast to conventional distillation methods, which often rely on clean images, Inverse Bridge Matching Distillation enables efficient knowledge transfer even with noisy data.

The effectiveness of Inverse Bridge Matching Distillation has been evaluated in extensive experiments. Various scenarios, including super-resolution, JPEG restoration, sketch-to-image transformation, and other tasks, were investigated. The results show that the distillation technique can accelerate the inference speed of DBMs by a factor of 4 to 100. In some cases, even an improved generation quality compared to the original teacher model could be achieved.

These advances in DBM distillation open up new possibilities for the use of diffusion models in practical applications. The faster inference speed allows the integration of DBMs into real-time systems and expands the range of applications to areas where fast image processing is essential. Inverse Bridge Matching Distillation thus represents an important contribution to the further development of diffusion models and their practical application.