Using LoRA for stabilization instead of EMA?

[jzhangbs]

Hi Hunyuan team, it's an awesome work! I have a question about using EMA for stabilizing the target model. After checking the released trainer of PCM, they use LoRA and do not use EMA. Will LoRA also achieve similar effect as EMA in 3D case?

[Zeqiang-Lai]

Hi @jzhangbs, I think training LoRA in text to image is for applying LoRA for different downstream models instead of stability. In 3D, we could also train a LoRA, so that a single distillation LoRA can be applied to Hunyuan3D-2 and Hunyuan3D-2mv without extra distillation.

Will LoRA also achieve similar effect as EMA in 3D case?
I am not sure, but it would be interested to make some investigation with experiments.