With the development of diffusion models, there are unique opportunities for the application of innovative methods, in particular for the generation of synthetic training datasets. However, there are a range of difficulties associated with computational resources and time when automating the augmentation of large amounts of data. We propose a new approach to generate training data in a similar domain as existing datasets. The proposed method is based on the Stable Diffusion v2, which has been further trained using a new loss function on a specially assembled TAOMR dataset in the street scene domain. Also we implement the automatic pipeline for generating fully synthetic datasets including the Depth Anything v2 synthetic depth model. The code for our method is publicly available: https://github.com/Areson251/GenLab