At present, a neural rendering model requires a long time to reconstruct a single scene, has high data requirements, supports only rigid objects and needs a long rendering time for the trained model. To solve these problems, different implementation versions of the classic neural radiance field (NeRF) model are analysed, and model training and rendering factors that affect its performance are explored and examined. Compared with the improved model of the typical NeRF, the performance analysis results of the neural radiation field technique are obtained. After identifying the major performance influencing factors of the neural radiation field, research ideas for the future improvement of this technology and its combined applications to different fields are provided.