An experimental voice style LoRA. Trained on an open dataset that includes both male and female and both normal dialogues and sex scenes.
Description
Better Japanese intonation
Tend to silence background sound
Lipsync may be broken
FAQ
Comments (13)
I can feel that it has improved significantly compared to Version 1. You did a great job.
I realized this while actively using voice prompt: when I don't use voice commands, moaning and slapping sounds come out simultaneously, but when I start using voice prompt, the slapping sound doesn't come out. that bums me out a little.Is this problem an issue with the LTX model itself?
@goaqjrj The dataset doesn't include any background noise, so I guess that's the reason. Lower weight may help
@2929dance The owner clearly stated in the description "Tend to silence background sound," but I didn't read it, sorry. I tried lowering lora strength to 0.9~0.5 as the owner suggested, but it didn't work as intended. Still, the voice comes through well, so I guess it doesn't matter. Anyway, thanks.
@goaqjrj I'll add normalization data in the next version. That should prevent overfitting to the silent background. Thanks for the feedback!
@2929dance It's literally just a horny fapper's greed, but thank you so much for reflecting it in the next version. I'll be looking forward to it.
Hi, could you create a LoRA for Wan2.2? : )
Unfortunately Wan 2.2 doesn't support sound
What tools do you use for training?
Adding a bit of this on anime generation is pretty nice. Thx
By the way, I am using Aitoolkit to train general Lora, but how do you focus on the sound ??
Thanks! I use this one: https://github.com/AkaneTendo25/musubi-tuner
@2929dance thanks will have a look !