Researchers from Sun Yat-sen University and other researchers proposed a new method called ScaleLong to address the stability issue during diffusion model training. This method effectively alleviates feature instability and enhances the model's robustness to input disturbances by performing scaling operations on the long skip connection of UNet. The researchers proposed two specific scaling coefficient adjustment methods: Learnable Scaling (LS) Method and Constant Scaling (CS) Method, and visually analyzed the role of features and parameters in the model training process, as well as the effect of the scaling coefficient on the gradient magnitude. and the impact of input disturbance stability. This research provides new ideas for improving the training stability and robustness of diffusion models.
The article focuses on:
Researchers from Sun Yat-sen University and other researchers proposed the ScaleLong diffusion model and pointed out that scaling operations on UNet's long skip connection can stabilize model training. Research has found that reasonably setting the scaling coefficient can alleviate feature instability and improve the model's robustness to input disturbances. They proposed Learnable Scaling (LS) Method and Constant Scaling (CS) Method, through which scaling coefficients can be adaptively adjusted to further stabilize model training. Visual features and parameters play an important role in the model training process, while the scaling coefficient affects the gradient magnitude and the stability of input perturbations.
The ScaleLong model effectively improves the stability and robustness of diffusion model training by improving UNet's long skip connection and combining Learnable Scaling and Constant Scaling methods, and provides important technical support for the application of diffusion models. Future research can further explore better scaling strategies to further improve model performance.