Different results in training vs inference

[A-2-H]

I am experimenting with training different models using different configs. Right now I am using Reflow from v2. But this problem occur in every config/models that I training with.

I noticed that the inference audio from my model has worse quality than validation file from training. Why is that? What can I do to achive quality from validation data?

input:
https://drive.google.com/file/d/1pMjKUtAgZwDqWL8A3YlaWV4B7I1SECYu/view?usp=sharing

val:
https://drive.google.com/file/d/1Byz-Zeg0kCm75hzmRxiXlwHA5MZp3R2O/view?usp=sharing

infer:
https://drive.google.com/file/d/1hh4dwiodXsESe_Rv7S49nMGEcl3pZ7lc/view?usp=sharing

[CNChTu]

The explanation of the configuration file will be provided after the experiment. Based on the current situation, using a 1024-width, 6-depth LynxNet network can already yield good results.
The better performance in the validation results could be due to this being an in-domain behavior.

[A-2-H]

So I need to use LynxNet instead of WaveNet?
By width and depth you mean n_layers and n_chans?

I am using reflow and by default in config_v2_reflow it is setup with Wavenet so how should I adjust it to achive configs for good results with Reflow? Or should I use different config file?