Epoch Demo Quality versus Normal Inference

[deccolquitt]

Hi,

I have been finetuning my own model with your finetuning colab notebook and the demos I am hearing on wanb are of much higher quality than when I try and generate audio from the same models through your 'dance diffusion' colab. Do you know what parameters the previews are generated with and how I can recreate them when generating audio from my models.

[zqevans]

To get the same setup in the Dance Diffusion notebook as the training demos, switch the sampler_type to v_ddim and set the step count to 250.

Let me know if this works.

[deccolquitt]

It's not quite worked - please listen to file attached. Both are from the same checkpoint, the first was generated through the DD notebook, and the second (starts on 22 seconds) was auto-generated whilst training. There is a lot more hiss and dead air in the background of the first.

Untitled.mov