Website fingerprinting enables attackers to snoop on users’ browsing preferences for websites, even if the network connections are encrypted. Fine-grained webpage fingerprinting can reveal more sensitive privacy by precisely distinguishing pages from the same website. Recognizing similar webpages within the same website needs representative fine-grained features. However, the fluctuating realistic network makes it challenging to ensure the stability of fine-grained features. In this paper, we propose a fine-grained webpage fingerprinting method named Stable Webpage Fingerprinting (StableWPF) to obtain stable webpage features that are not affected by the unstable realistic network. We use the frequency information of the length of the TLS fragments to depict the most representative features. To eliminate the fluctuation of the features, we leverage the kernel density estimation and the Bag-of-Words model in our method. The experimental results in the closed-world and open-world scenarios show that our method outperforms the state-of-the-art approaches in accuracy, precision, and recall. The robust performance obtained on famous real-world websites with various network environments demonstrates the generalization ability of our method.