Using Stable Audio 3 to generate variations of an existing loop. Unconditional generation (no prompt), renoising the latents to 0.5, and just using different seeds seems to generate a nice neighbourhood around the original. Generally keeps the harmonic context and feel.

Cool! Additionally, with Stable Audio 3, its fast inference speed feels like it could enable some very interesting possibilities in live music.

Yeah! Small models usable on device in real time, big unlock

crazy

Impressive.

Any links to project?

alas not to my specific setup, but all the interesting stuff is done by the stable audio 3 medium model

I wonder how well it will work for time-stretching. This is the most painful point in my sampling flow currently, old tracks just cannot be stretched into the grid with the common algorithms

Good question. I don't have an intuition for how stretched latents might decode. Could be an interesting thing to try. I guess the inpainting stuff they have is more ready at hand, where you can insert new parts to a sample rather than stretch.

what's the original music?

this loop from splice

thanks 🫡

This is really dope, what's UI? Is it public?

thx! the ui is a hastily vibe coded remix of this

Nice! Should also be possible to "inpaint" between the samples to create transitions?

Absolutely! Need to experiment with how much inpainting remains on course without prompts, or whether it needs guidance

amazing