About the Guidance Scale

[robintzeng]

Hi, I have a question about the guidance scale.

In the repository, when generating the ODE pairs, the guidance scale is set to the default value of 6.0. However, during DMD training, the guidance_scale is set to 3.0 in the YAML file.

How do you decide on these values? And is there a specific reason why the guidance scale used for generating ODE pairs is larger than the one used for DMD training?

Thanks!

[zhuhz22]

This has not been studied in detail with ablation experiments. The CFG scale at different stages was directly referenced from CausVid, which may indeed not be optimal. A detailed investigation will be conducted in the future.