Nowadays, the demand for photorealistic rendering in the movie, anime, game, and other industries is increasing, and the highly realistic rendering of 3D scenes usually requires a lot of calculation time and storage to calculate global illumination. How to ensure the quality of rendering on the premise of improving drawing speed is still one of the core and hot issues in the field of graphics. The data-driven machine learning method has opened up a new approach. In recent years, researchers have mapped a variety of highly realistic rendering methods to machine learning problems, thereby greatly reducing the computational cost. This article summarizes and analyzes the research progress of highly realistic rendering methods based on machine learning in recent years, including: global illumination optimization calculation methods based on machine learning, physical material modeling methods based on deep learning, and participatory media drawing method optimization based on deep learning, Monte Carlo denoising method based on machine learning, etc. This article discusses the mapping ideas of various drawing methods and machine learning methods in detail, summarizes the construction methods of network models and training data sets, and conducts comparative analysis on drawing quality, drawing time, network capabilities, and other aspects. Finally, this article proposes possible ideas and future prospects for the combination of machine learning and realistic rendering.