Blunstron is a model I made for Harmonai's Dance Diffusion.
The dataset is less than five minutes of the song Old and Wise by The Alan Parsons Project, yet it performs very well and does not overfit.
Old and Wise is sung by Colin Blunstone, hence the name Blunstron.
Why
I put music out for free on YouTube containing lots of tiny samples (in imitation of a musician named Todd Edwards),
and I've sampled Old and Wise a LOT because I like the auditory textures in it. I'm kind of running out of potential chops,
so I decided to generate a practically infinite supply of them with AI.
How
I finetuned this on Google Colab for around two hours. I kind of dislike how the word "finetune" is used for Dance Diffusion, since unlike
Dreambooth with Stable Diffusion, Dance Diffusion models (including this one) effectively
become an entirely different model when fine-tuned for long enough.
Audio Characteristics
- Deep, ethereal, soft auditory texture
- Mostly chords, not much melody
- Colin Blunstone-like vocals
- Occasional drum hits
A few examples are provided in the files of this git repo called blunstron-test-x.wav.
