Commit
•
abf7d9d
1
Parent(s):
6ded603
Update README.md
Browse files
README.md
CHANGED
|
@@ -6,6 +6,8 @@ https://twitter.com/_lyraaaa_/status/1819145905972691227
|
|
| 6 |
|
| 7 |
finetuned stable audio open's vae for 100k steps to try and fix its habit of colorizing gritty sounds
|
| 8 |
|
|
|
|
|
|
|
| 9 |
the blue and orange runs are near-identical, same seed etc, except the orange one had the encoder and bottleneck frozen while blue was a full train. orange model has an identical latent space and therefore is instantly swappable into any stable audio open model, blue will require further training in exchange for slightly higher fidelity.
|
| 10 |
|
| 11 |
to use the blue vae, pass it to your train command with --pretransform-ckpt-path.
|
|
|
|
| 6 |
|
| 7 |
finetuned stable audio open's vae for 100k steps to try and fix its habit of colorizing gritty sounds
|
| 8 |
|
| 9 |
+
|
| 10 |
+
|
| 11 |
the blue and orange runs are near-identical, same seed etc, except the orange one had the encoder and bottleneck frozen while blue was a full train. orange model has an identical latent space and therefore is instantly swappable into any stable audio open model, blue will require further training in exchange for slightly higher fidelity.
|
| 12 |
|
| 13 |
to use the blue vae, pass it to your train command with --pretransform-ckpt-path.
|