In the early days of unsupervised image-to-image translation, models like CycleGAN and UNIT revolutionized the field by learning to map one domain (e.g., horses) to another (e.g., zebras) without paired examples. However, these models suffered from a "combinatorial explosion." If you wanted to translate between $N$ different domains, you traditionally needed to train $N(N-1)$ separate models. For ten domains, that is 90 models; for twenty, it is 380.
A growing body of research has shifted toward a more scalable architecture: mapping all domains into a shared latent space. By decoupling domain-specific features from shared structural information, these researchers have reduced the computational complexity of multi-domain translation from quadratic ($O(N^2)$) to linear ($O(N)$).
The Hub-and-Spoke Intuition
The core idea is to stop thinking of translation as a direct bridge between two specific points. Instead, think of it as a "hub-and-spoke" system. Every domain has its own "spoke" (an encoder and a decoder) that connects to a central "hub" (a shared latent representation).
To translate an image from Domain A to Domain B, the system first encodes the image into the universal hub representation and then decodes it using Domain B’s specific decoder. Asha Anoosheh and Eirikur Agustsson formalized this in ComboGAN, demonstrating that by assigning just one encoder and one decoder to each domain, the total number of sub-networks scales linearly. This allows researchers to scale to dozens of domains with a fraction of the hardware resources previously required.
Decoupling Content from Style
For this hub-and-spoke model to work, the latent space must be "domain-agnostic." It should capture the underlying structure (the "content") while stripping away the domain-specific "style."
Le Hui and Xiang Li explored this by decoupling domain-specific features from shared structural information. Similarly, Jinsung Yoon and Mihaela van der Schaar introduced RadialGAN, which uses this central latent space to reconcile mismatched feature sets from disparate datasets. This architecture ensures that even if datasets have different properties, they can still contribute to a unified model.
Modularity and the "Plug-and-Play" Domain
One of the most powerful consequences of this linear scaling is modularity. In a quadratic system, adding an 11th domain requires training 20 new bridges to connect it to the existing 10. In a linear system, you only need to train one new encoder-decoder pair.
Caroline Uhler and Karren Yang demonstrated that learning uncoupled autoencoders allows for the addition of new data domains without retraining or even accessing the data of previously integrated domains. This "plug-and-play" capability is vital for data privacy and long-term system maintenance.
This modularity has been extended to diffusion models by Jiaming Song and Stefano Ermon. Their Dual Diffusion Implicit Bridges (DDIBs) treat translation as a sequence of two independent diffusion steps. Because each domain is modeled independently, the system achieves exact cycle consistency and scales linearly without ever requiring joint training on multiple domains simultaneously.
Harmonization in Heterogeneous Environments
The practical impact of this research is most visible in medical imaging, where "domains" aren't just artistic styles, but different scanner manufacturers or hardware settings.
- OCT Scans: Gaurav Fotedar and Ziheng Zhou developed MDT-Net, which uses domain-specific modules to harmonize scans across different devices. This ensures that a diagnostic AI performs consistently regardless of which machine took the image.
- CT Reconstruction: Aravind R. Krishnan and Thomas Z. Li applied multipath cycleGANs to standardize noise and resolution across various scanner vendors.
By mapping these heterogeneous hardware outputs into a shared latent space, researchers like Ziliang Chen have shown that we can match joint distributions across many domains efficiently, moving past the limitations of pairwise models toward truly universal translation frameworks.