FR: Generate forever, output directory and wildcards

[Arthur-McArthur]

Would be useful to have an easily accessible output folder.

Generate forever would help us make many outputs and cherry-pick the best ones.

Wildcards could help create unusual and unexpected outputs.

[Duemellon]

yes please. Saving the last output location OR allowing us to set the output directory would be great.

[noprompt]

You can easily adapt the notebook code, your own scripts, etc. to export the audio:

from scipy.io.wavfile import write as write_wav
import uuid

descriptions = [
    'lofi 80s workout music',
    'downtempo vhs vibes',
]

output = model.generate(
    descriptions=descriptions,
    progress=True
)

for i, description in enumerate(descriptions):
  file_id = uuid.uuid4()
  # Write a text file containing the description.
  open(f"{file_id}.txt", "w").write(description)
  # Write the wav file
  write_wav(f"{file_id}.wav", 32000, output[i].squeeze().cpu().numpy()

[Duemellon]

You can easily adapt the notebook code, your own scripts, etc. to export the audio:

I appreciate the guidance/support. Unfortunately I'm one of those 3rd gen interface users -- not the designer (1st gen), nor on level with coding (2nd gen), but that 3rd gen that is right above that group that can only use it if it's a one-click install.

I run it locally via Gradio (preferred). What file do I add this to? Where?

[JonathanFly]

I've got the other stuff done, I just keep getting distracted and never pushed a fork public. It sounds like you got a very specific and complicated work flow in mind? I always have trouble imaging people who aren't me using the UI, can you elaborate? Or if it's too off topic for an issue on a git, send me a 'dm' via some channel. It would help to know how people use things, before I build..

[Arthur-McArthur]

I am envisioning a workflow similar to how I use Stable Diffusion locally, where I would write prompt syntax like:

" a {dark|mysterious|happy} melody played by a {flute|guitar|trumpet} with a {fast|slow} tempo "

and then hit generate forever and let it create many different variations within the output folder, named with the prompt. For example the name of an item could be: "00001-a_dark_melody_played_by_a_guitar_with_a_slow_tempo".

Stable Diffusion has a "Dynamic Prompt" extension that allows for reading wildcards out of text files as well.

[StableInfo]

Hello @Arthur-McArthur , were you able to impement it? @Duemellon did you find which file had to be changed? I think it can be done easily using something like notepad++ the opening one of the files. Thanks @noprompt still waiting for more guidance, is it inthe musicgen_app file?

[StableInfo]

@Duemellon , open musigen_app with notepad++, then go to the line around 150 ish, "print("Tempfiles currently stored: ", len(file_cleaner.files))", at the end of it press enter to insert an empty line of code, and paste there, this:

file_id = uuid.uuid4()
write_wav(f"{texts[0]}.wav", 32000, outputs.squeeze().cpu().numpy())

Thats it. Make sure every line is alined with the lines above and before it. Now restart the app with python .\demos\musicgen_app.py --inbrowser if its already open the clode it with CTRL+C Now every file is saved inside the directory and named like our prompt.

Make sure to add

from scipy.io.wavfile import write as write_wav
import uuid

At the beginning of the file somewhere. Enjoy (be careful though, if you have the same prompt and different parameters, new files will delete old ones)