proper forge
download: https://huggingface.co/lyery/models/resolve/main/galacta_257e_22616s.rar?download=true
BETTER STOP THAT VEHICLE
well, you can’t expect to win them all 🤪
index 0 to 0.5 sounds fine
model info 🤓 :
dataset: 15 minutes
batch size: 4
f0: rmvpe
sample rate: 32k
pretrain: klm 4 final
vocoder: hifigan
steps: 22.616
This model is compatible with mainline, older forks and w-okada voice changer
river oracle
My fx take on galacta
proper forge
My fx take on galacta
this is way more accurate than my attempt :0 good job
river oracle
i thought she was going to be much harder ngl, but shes just reverb and delay
proper forge
i used echo and reverb
river oracle
fickle forum
How do you detect overtraining properly, im gonna go insane 😭
Its so perfect..
proper forge
How do you detect overtraining properly, im gonna go insane 😭
Its so perfect..
its pretty easy
compare lets say epoch 200 vs epoch 500, if epoch 500 dont generate any frequencies in parts where epoch 200 did generate them, you can assume epoch 500 is already overtrained
basically the model forgets the pretrain and can only inference stuff that is very close to the dataset
and sometimes overtrained models can only inference their own dataset and nothing more
sounds confusing but trust me, hearing overtraining is very easy, the model just starts to sound very robotic
example:
epoch 200 sounds fine, has ok singing and the model is capable of generating frequencies, this epoch not overtrained
epoch 500 now sounds robotic/distorted and the model struggles to do anything so bad that it doesnt know how to clone frequencies anymore, everything now sounds worse compared to epoch 200 results