Dense forecasting is a method that aims to predict future representations by utilizing information from previous images. It encompasses a range of techniques that find applicability in various domains such as autonomous vehicles and video prediction. This paper explores different approaches to dense forecasting by leveraging modified pyramidal SwiftNet architecture as semantic feature extractor and utilizing VQ-GAN for extracting latent representations used for reconstruction. We evaluate our experiments on the Cityscapes dataset. The forecasting process utilizes existing F2F and F2MF architectures, and introduces several model modifications with the objective of enhancing prediction accuracy. Particular attention is given to assessingthe impact of these modifications on inference speed and computational complexity. Notably, the inclusion of higher-resolution feature forecasting yields a two percentage-point improvement in the mIoU metric for both models, while the mIoU-MO metric, focusing exclusively on moving objects, exhibits a three percentage-point increase. Furthermore, a multimodal architecture is proposed for joint forecasting of semantic and reconstruction representations. The forecasting of VQ-GAN reconstructions results in blurring of in-motion sections of the scene, while achieving accurate and precise reconstructions in static sections.