Analysing tissue sections is crucial in cancer diagnosis, influencing tumour classification and treatment decisions. Hematoxylin and Eosin (HE) staining is widely used; however, it captures only basic morphological structures, prompting pathologists to use Immunohistochemistry (IHC) for more detailed information, such as the Ki67 index, which indicates cell proliferation. However, IHC is more time- and resource-consuming. Deep learning models offer potential enhancements in medical diagnosis by providing consistent and cost-effective decisions. For clinical application, these models must be explainable. A generative model can provide additional information that can help to explain predictions. This paper presents a conditional StyleGAN model to generate synthetic HE-stained images conditioned on the Ki67 index. We evaluate three StyleGAN models: one trained on an unfiltered dataset, another on a filtered dataset that includes only higher-quality HE-stained images with a sufficient number of cells, and a third model trained on the filtered dataset with extended training duration. We analyse the results in terms of training progress and the quality of generated images. Results indicate that models trained on the filtered dataset generate more realistic images, and the model with the extended training duration produces indistinguishable images from real samples validated by the expert pathologist.