Reference-based super-resolution (RefSR) has gained attention for its superior performance due to the introduction of high-quality external priors. However, existing RefSR methods all rely on paired images and require manual selection of reference images in practical applications. To address these challenges, this paper proposes a framework for generating high-quality reference images, thereby overcoming the difficulties associated with manual selection in RefSR. We propose a novel Self-Supervised Reference-based Image Super-Resolution method (SSR-SR), which employs a conditional diffusion model and self-supervised learning (SSL) representations to generate reference images with a high degree of semantic similarity to the input image. Since reference images can prioritize perceptual quality over fidelity, we enhance these reference images using a diffusion-based super-resolution approach. The framework also includes a dynamic aggregation module and a contrastive alignments network to ensure precise texture transfer and robust alignment between the low-resolution (LR) input and the high-resolution (HR) reference. Experimental results on multiple benchmarks demonstrate that our proposed SSR-SR achieves competitive results without relying on paired data. This work highlights the potential of diffusion models and SSL representations in advancing the field of image super-resolution.