Along with advances in biology and the simulation of nano electronic devices, atomic structures play a crucial role in modern science and technology. The complex details of the atomic structure result in the far-reaching impact of the position of the light source on the rendering effect, incurring difficulties in rendering atomic models. On this basis, an atomic model rendering method based on a reference image was proposed, in which the lighting parameters of the reference image were calculated for the rendering of the atomic model. First, a POV-Ray script was used to render a batch of models at different light angles by changing the light source positions, and the light source position parameters and rendered images were collected to obtain a dataset of rendered images corresponding to the light source positions. Then, the light source estimation network was designed with the residual neural network as the backbone, and the attention mechanism was embedded in the network to enhance the network accuracy. The optimized light source estimation network was employed to train the dataset and regress the light source location parameters. Finally, the trained convolutional neural network was used to estimate the rendering parameters of the reference image, and the target model was rendered using the rendering parameters. The experimental results show that the parameters predicted by the network are highly reliable with minimal error compared with the real lighting parameters.