Ars Technica
On Thursday, a pair of tech hobbyists launched Riffusion, an AI mannequin that generates music from textual content prompts by creating a visible illustration of sound and changing it to audio for playback. It makes use of a fine-tuned model of the Stable Diffusion 1.5 picture synthesis mannequin, making use of visible latent diffusion to sound processing in a novel manner.
Created as a passion undertaking by Seth Forsgren and Hayk Martiros, Riffusion works by producing sonograms, which retailer audio in a two-dimensional picture. In a sonogram, the X-axis represents time (the order wherein the frequencies get performed, from left to proper), and the Y-axis represents the frequency of the sounds. In the meantime, the colour of every pixel within the picture represents the amplitude of the sound at that given second in time.
Since a sonogram is a sort of image, Steady Diffusion can course of it. Forsgren and Martiros skilled a customized Steady Diffusion mannequin with instance sonograms linked to descriptions of the sounds or musical genres they represented. With that information, Riffusion can generate new music on the fly based mostly on textual content prompts that describe the kind of music or sound you need to hear, equivalent to “jazz,” “rock,” and even typing on a keyboard.
After producing the sonogram picture, Riffusion makes use of Torchaudio to vary the sonogram to sound, taking part in it again as audio.
“That is the v1.5 Steady Diffusion mannequin with no modifications, simply fine-tuned on photos of spectrograms paired with textual content,” write Riffusion’s creators on its explanation page. “It may well generate infinite variations of a immediate by various the seed. All the identical net UIs and methods like img2img, inpainting, damaging prompts, and interpolation work out of the field.”
Guests to the Riffusion web site can experiment with the AI model due to an interactive net app that generates interpolated sonograms (easily stitched collectively for uninterrupted playback) in actual time whereas visualizing the spectrogram repeatedly on the left facet of the web page.
It may well fuse kinds, too. For instance, typing in “clean tropical dance jazz” brings in components of various genres for a novel consequence, encouraging experimentation by mixing kinds.
In fact, Riffusion is just not the primary AI-powered music generator. Earlier this yr, Harmonai launched Dance Diffusion, an AI-powered generative music mannequin. OpenAI’s Jukebox, introduced in 2020, additionally generates new music with a neural community. And web sites like Soundraw create music continuous on the fly.
In comparison with these extra streamlined AI music efforts, Riffusion feels extra just like the passion undertaking it’s. The music it generates ranges from attention-grabbing to unintelligible, however it stays a notable software of latent diffusion expertise that manipulates audio in a visible area.
The Riffusion mannequin checkpoint and code are available on GitHub.
