Question about the runtime

[zonszer]

Hi,

Thank you for your excellent work. I have a question about NeoVerse’s runtime. From my understanding, you use a distillation-based approach to reduce the number of inference steps at test time. Could you share more details about how this is done?

For example, did you use LightX2V (or a similar method) to distill the model after the original training, train a LoRA for few-step inference, and then apply that LoRA to both the main branch and the control/context branch?

Thanks for your time!

[YuxueYang1204]

Hi @zonszer ! We adopt lightx2v/Wan2.1-T2V-14B-StepDistill-CfgDistill-Lightx2v to reduce inference steps. The LoRA is directly applied to the main branch since it remains frozen during the training.

Hope this helps!