Prompt Tuning Is All We Need?

Abstract

Recent advances in pre-trained vision-language models, e.g., CLIP, have demonstrated remarkable success in domain generalization (DG) by tuning prompts. To promote DG, one promising method is to explore how to design or learn more sweet prompts, i.e., prompt learning. The implicit intuition of it is that a more elaborate prompt learning method can lead to higher generalization performance. The foundation intuition motivates us to raise a question: Prompt tuning is all we need? To verify whether the intuition holds for DG, we design comprehensive experiments on DG benchmarks. However, our experiments demonstrate a pessimistic conclusion that simply tuning prompts using training sets can achieve comparable performance with that using test sets. Namely, even the optimal prompts can hardly bring significant performance gain than a simple tuning strategy. Our experiments show that this results from the non-separability of features extracted by the image encoder. Thus, we propose image encoder tuning, named Im-Tuning, for more separable image features. We conduct extensive experiments on multiple DG benchmarks, demonstrating that Im-Tuning can consistently outperform the relevant state-of-the-art methods.