This paper presents a workflow for estimating the 3D morphological characteristics of latex aggregates from 2D in-situ images using deep learning and stochastic geometry models. The method includes automatic image segmentation using a Convolutional Neural Network (CNN), 3D object generation using a Generative Adversarial Network (GAN), and estimation of 3D characteristics. Validation with synthetic datasets shows effective size, shape, and texture characterization, with the Mean Absolute Percentage Error (MAPE) for morphological characteristics of generated objects being around 5% at most. Application to real in-situ images demonstrates feasibility and consistency with experimental observations, successfully generating a digital twin of the latex aggregate population. The method’s flexibility and efficiency make it suitable for real-time industrial applications, offering potential for process monitoring and quality control. Future work will focus on enhancing model performance and adapting to different particle types for broader applicability in various industrial settings.