Hugging Face's logo

NekoMikoReimu
/
TohoWorkInProgress

TensorBoard
Model card Files Metrics Community
TohoWorkInProgress / README.md
NekoMikoReimu's picture
NekoMikoReimu
Update README.md
c4ab161
preview code
|
raw
history blame contribute delete
844 Bytes

Highly Responsive to Prompts - Touhou MusicGen(Medium) Finetune Notes

11/13 Research Notes:

  1. Some of the generated samples are going in ways that sound not-touhou-ish, but more traditional upbeat/cheerful, like elevator or corporate filler music. A few potential causes: a. The Essentia-generated audio tags haven't been looked at carefully to see if they actually match the song in question, I remember seeing some wackiness in there that we could stand to clear out..
    b. I'm not sure, but I think I heard some key changes in the samples. Should check the "key" items in the jsonl as well.
  2. For our next run, I'd like to add in LpMusicCaps support. https://huggingface.co/spaces/seungheondoh/LP-Music-Caps-demo
  3. I'm not sure that the downgrading to 32khz is actually necessary, I've seen other people not do it.