the training setting of audio diffusion branch

[mabingqi]

Thanks for the great open-source work. I have a few questions about training the audio diffusion branch:

1. Could you share more details about the training setup—such as batch size, number of training iterations, and the timestep sampling strategy?
2. I’ve tried training an audio diffusion model on top of Hunyuan-Foley, but I often observe artifacts such as electrical noise, and the convergence is much slower than adopting mmaudio. I’m not sure whether you’ve encountered similar behavior.

Looking forward to your reply.