Ask HN: How is Stable Diffusion different from sampling in hip-hop?

https://news.ycombinator.com/item?id=33733679

by shanebellone • 4 years ago

2 2 4 years ago

news.ycombinator.com

Stable Diffusion looks suspiciously like hip-hop sampling in the 80s.

How do these products really differ? Will Stable Diffusion face the same scrutiny? Could licensing be a solution (think official training sets)?

jonas_kgomo• 4 years ago

Sampling in hip hop is about taking 1 or 2 audio sources mapping them into some new audio, SD was trained on 2.3 billion data points. In a diffusion model you put noise on an input image, until its recognizeable. For a song this would mean, making it unrecognizable by making it literally sound like noise, and then try to denoise it, this process wouldn't be very helpful i assume, but might be a good experiment. Open AI's Jukebox uses an autoenconder method with autoregressive transformers, given this architecture, i would assume you can theoretically do stable diffusion architecture on audio, to generate new audio. https://openai.com/blog/jukebox/

shanebellone• 4 years ago

The processes are essentially identical while the specifics and quantities differ.

- acquire samples

- process until output is "recognizable"

- manipulate until output is desirable

Stable Diffusion requires IP as an input.

Ask HN: How is Stable Diffusion different from sampling in hip-hop?

Related Stories