Prompt Disentanglement via Language Guidance and Representation Alignment for Domain Generalization
Researchers have developed a new framework for domain generalization in computer vision that leverages language guidance from pre-trained Visual Foundation Models (VFMs). The method first disentangles text prompts using a large language model (LLM) and then uses these disentangled text features to guide the learning of domain-invariant visual representations. To further enhance robustness, an additional component called Worst Explicit Representation Alignment (WERA) is introduced, which uses abstract prompts and stylized image augmentations to ensure consistency across different visual distributions. Experiments on several benchmark datasets show that this approach surpasses existing state-of-the-art domain generalization techniques. AI
IMPACT This research could lead to more robust AI models that perform better across different, unseen datasets without requiring retraining.