Image generation models aim to generate realistic images according to certain requirements. In recent years, there has been significant progress in the application of machine learning in computer vision. Various application models are becoming increasingly complex in order to improve model performance, resulting in increasingly intricate loss functions and a growing number of hyperparameters for weighting these loss functions, leading to high training costs for the models. We propose a method for stepwise training of models, called SplitObject, which splits the intricate loss function to reduce training difficulties, reduce the number of hyperparameters, and simplify the training process of the model. For loss functions composed of multiple terms, we split the loss function into multiple parts based on the functions of each term, and improve the convergence speed of the model by gradually adding each part during training. For a specific term in the loss function, we divide it into several blocks to focus the model’s training on a specific aspect of the loss function during certain periods. We also apply this method to training other machine learning tasks. Our experiments show that our method can make the model converge faster and simplify the training process. Our experimental results demonstrate that models trained using our SplitObject method generate images with superior FID scores.