This project aims to find ways to improve the ability of Diffusion Models to generate 3D consistent novel views of an object or scene. Current methods are limited in the geometric consistency of generated views, and the fine-tuning process is very heavy (taking up to 7 days using  8xA100 GPUs). We aim to investigate the latent space of these diffusion models and to find efficient ways of utilizing geometric constraints to improve consistency, focusing specifically on improving results for scenes and not just single objects. We also want to investigate how video diffusion models can be used for this.