Universal Multi-Domain Translation via Diffusion Routers

2025-09-26

arXiv.org

AI Summary

Multi-domain translation (MDT) is fundamentally limited by its reliance on fully aligned multi-domain paired data or support only for seen domain pairs, hindering generalization to arbitrary K-domain translation. This paper proposes Universal Multi-Domain Translation (UMDT), enabling bidirectional translation among any K domains using only K−1 paired datasets—each sharing a common central domain. Methodologically, we introduce a diffusion-based router that jointly models direct mappings and indirect mappings via the central domain, coupled with a variational bound learning strategy. We employ a single noise predictor conditioned on domain labels, enhanced by Tweedie refinement and scalable training. To our knowledge, UMDT is the first approach achieving universal MDT without requiring fully aligned multi-domain data. It establishes new state-of-the-art performance on three major benchmarks, reduces sampling cost, and successfully extends to novel tasks such as sketch ↔ semantic segmentation translation.

Technology Category

• Machine Learning: Transfer, Domain Adaptation, Multi-Task Learning
• Natural Language Processing: Machine Translation, Multilinguality, Cross-Lingual NLP
• Application Domains: Transportation

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Web Mining and Content Analysis: Mining multimedia, multimodal, multilingual, cross-lingual Web data
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Multi-domain translation (MDT) aims to learn translations between multiple domains, yet existing approaches either require fully aligned tuples or can only handle domain pairs seen in training, limiting their practicality and excluding many cross-domain mappings. We introduce universal MDT (UMDT), a generalization of MDT that seeks to translate between any pair of $K$ domains using only $K-1$ paired datasets with a central domain. To tackle this problem, we propose Diffusion Router (DR), a unified diffusion-based framework that models all central$leftrightarrow$non-central translations with a single noise predictor conditioned on the source and target domain labels. DR enables indirect non-central translations by routing through the central domain. We further introduce a novel scalable learning strategy with a variational-bound objective and an efficient Tweedie refinement procedure to support direct non-central mappings. Through evaluation on three large-scale UMDT benchmarks, DR achieves state-of-the-art results for both indirect and direct translations, while lowering sampling cost and unlocking novel tasks such as sketch$leftrightarrow$segmentation. These results establish DR as a scalable and versatile framework for universal translation across multiple domains.

Problem

• Enables translation between any domain pair using limited paired data
• Models all translations with a single diffusion framework
• Supports both indirect and direct non-central mappings efficiently

Innovation

• Diffusion Router framework for multi-domain translation
• Scalable learning with variational-bound objective
• Tweedie refinement for efficient direct mappings