Lora?

[Oruli]

Does this require loras to be retrained? Because with current loras the motion just doesn't work at all.

[aejion]

Unfortunately, since our method involves finetuning all parameters of HunyuanVideo, the LoRA may also need to be retrained.

[Oruli]

Unfortunately, since our method involves finetuning all parameters of HunyuanVideo, the LoRA may also need to be retrained.

It would be nice to get some certainty on that, maybe document it also? There are 1000s of loras for hun, nobody is going to retrain everything, especially if they don't even know they have to.

[aejion]

You can try using the quantized version (follow this link). It seems that AccVideo supports using pre-trained LoRAs.