The invention discloses a label construction method, system and device for a training model, and a storage medium, and the method comprises the steps: carrying out the label extraction of a figure image, simplifying a complex figure image into a key point two-dimensional coordinate or a two-dimensional mask, and carrying out the training of a generative adversarial neural network model (GAN model); different label images can be generated by simply modifying coordinate positions or two-dimensional mask shapes of key points, figure images corresponding to labels can be generated by inputting thetrained generative adversarial neural network model, and then videos are further synthesized, so that the operation process of figure video synthesis is greatly simplified, and the working efficiencyis improved; subsequently, a new label can be added as a limiting condition according to actual generation requirements, the label and the real image corresponding to the label are sent to the generative adversarial network model for training, and finally, the corresponding real image can be generated through expansion conditions. The method is widely applied to the technical field of image processing.