Reproduce “Noise Optimized Conditional Diffusion for Domain Adaptation” on OpenResearch

About this paper

Lingkun Luo, Shiqiang Hu, Liming Chen

A new framework, Noise Optimized Conditional Diffusion for Domain Adaptation (NOCDDA), unifies conditional diffusion models with domain adaptation classifier training. It introduces a class-aware noise optimization strategy to generate highly discriminative target domain samples, leading to superior performance with an average accuracy of 99.3% on digital image benchmarks and 89.4% on office object recognition tasks.

Key insights

• The quality and discriminative utility of pseudo-labeled target domain samples are paramount for Unsupervised Domain Adaptation, significantly more impactful than just increasing their quantity.
• Integrating conditional diffusion models with DA classifier training through a task-coupled optimization, where a classifier is robust to noise from forward diffusion and guides reverse generation, is highly effective.
• Replacing standard Gaussian noise initialization with class-specific and classifier-guided noise optimization during reverse diffusion is critical to generate discriminative, rather than "class-confused," target domain samples.

Summary by alphaXiv.