|
|
--- |
|
|
license: mit |
|
|
base_model: |
|
|
- kronosta/blunstron |
|
|
library_name: diffusers |
|
|
--- |
|
|
|
|
|
This is another dance diffusion AI similar to Blunstron, except this time it generates harmonizing voices, and sings in either E or G half sharp. |
|
|
I finetuned it from Blunstron on a dataset of 3-second clips of Jacob Collier - In The Bleak Midwinter (which like Blunstron's dataset of The Alan Parsons Project - |
|
|
Old and Wise, I have sampled dozens of times in my "audio collage" music). However the samples were too quiet for my tastes and always sung in |
|
|
F, because it latched on to the beginning of the song. So I halted the training, removed all the quieter parts from the dataset, then resumed from where I left off |
|
|
so it could keep the knowledge it got already. This time I decided to preserve the entire history of checkpoints, since each once is unique. |
|
|
|
|
|
# Checkpoints |
|
|
- `epoch=948-step=192000.ckpt` is probably the best one, and it's the one I'll probably use. It sings in E and G half sharp, is varied from the original dataset, |
|
|
and is sometimes composed and sometimes wild, for the variety. |
|
|
- `epoch=1019-step=192500.ckpt` seems to be overfit, since a lot of the stuff it generates I can immediately hear where in the song it came from, from memory alone. |
|
|
Its perfect sound doesn't leave much room for creativity and also definitely qualifies as art theft. Like the previous, it sings in E and G half sharp. |
|
|
- `epoch=876-step=191500.ckpt` is from before the deletion of most of the dataset. It sings in F and is relatively quiet and composed. |
|
|
- `epoch=845-step=191000.ckpt` I have no training samples from, and haven't tested. It's quite likely it might share some aspects with Blunstron. |
|
|
|
