There is a close relationship between the size and property of a reservoir and the production and capacity. Therefore, in the process of oil and gas field exploration and development, it is of great importance to study the macro distribution of oil–gas reservoirs, the inner structure, the distribution of reservoir parameters, and the dynamic variation of reservoir characteristics. A reservoir model is an important bridge between first-hand geologic data and other results such as ground stress models and fracture models, and the quality of the model can influence the evaluation of the sweet spots, the deployment of a horizontal well, and the optimization of the well network. Reservoir facies modeling and physical parameter modeling are the key points in reservoir characterization and modeling. Deep learning, as an artificial intelligence method, has been shown to be a powerful tool in many fields, such as data fusion, feature extraction, pattern recognition, and nonlinear fitting. Thus, deep learning can be used to characterize the reservoir features in 3D space. In recent years, there have been increasing attempts to apply deep learning in the oil and gas industry, and many scholars have made attempts in logging interpretation, seismic processing and interpretation, geological modeling, and petroleum engineering. Traditional training image construction methods have drawbacks such as low construction efficiency and limited types of sedimentary facies. For this purpose, some of the problems of the current reservoir facies modeling are solved in this paper. This study constructs a method that can quickly generate multiple types of sedimentary facies training images based on deep learning. Based on the features and merits of all kinds of deep learning methods, this paper makes some improvements and optimizations to the conventional reservoir facies modeling. The main outcomes of this thesis are as follows: (a) the construction of a training image library for reservoir facies modeling is realized. (b) the concept model of the typical sedimentary facies domain is used as a key constraint in the training image library. In order to construct a conditional convolutional adversarial network model, One-Hot and Distributed Representation is used to label the dataset. (c) The method is verified and tested with typical sedimentary facies types such as fluvial and delta. The results show that this method can generate six kinds of non-homogeneous and homogeneous training images that are almost identical to the target sedimentary facies in terms of generation quality. In terms of generating result formats, compared to the cDCGAN training image generation method, traditional methods took 31.5 and 9 times longer. In terms of generating result formats, cDCGAN can generate more formats than traditional methods. Furthermore, the method can store and rapidly generate the training image library of the typical sedimentary facies model of various types and styles in terms of generation efficiency.