Why do you not use validation on SAD fine tuning

[Ifrah-Arif]

I am finetuning SAD model and it only provides training loss metric, why did you not use validation in sad class?

[AaronZ345]

No, I intentionally did not add a full validation loop for SAD.
In this project, SAD is a post-refinement stage, and the practical validation signal is weak relative to the cost:

1. Evaluation needs full synthesis for diffusion, which is expensive and slow.
2. A simple val-loss log is not very informative for model selection in our setup.