Translated data: Researchers from Sun Yat-sen University and other institutions have proposed the ScaleLong diffusion model, which indicates that the scaling operation on the long skip connection of UNet can stabilize model training. Their studies have found that setting the scaling coefficient appropriately can alleviate feature instability and enhance the model's robustness to input perturbations. They introduced the Learnable Scaling (LS) Method and the Constant Scaling (CS) Method, which allow for the adaptive adjustment of scaling coefficients to further stabilize model training. Visualization of features and parameters plays a crucial role during the training process, while the scaling coefficient influences the magnitude of gradients and the stability against input disturbances.