Lines are notes
The wavering lines in “Grandma’s Porch” are single held notes. Draw a line in the wrong place and the pitch goes with it.
Riffusion: music from a picture
A 2022 experiment that made music by drawing it. Type a prompt in the community demo below and see the picture along with the sound.
Riffusion, by Seth Forsgren and Hayk Martiros, adapts an image model to draw spectrograms, which are pictures of sound, and then turns each picture into audio. The original code is no longer maintained. This community demo makes 5 to 10 seconds.
The project’s own demo was showing an error when we checked, so this one is a community copy on Hugging Face. It may queue or stop, spends no credits, and we don’t run it.
Open on Hugging Face- See the sound as a picture
- 5 to 10 seconds
- A piece of music history
How to read a spectrogram
Riffusion’s whole trick depends on these pictures, so it helps to read one. This is all of “Neon at 2AM”, synthwave from our own generator, drawn by our tools rather than by Riffusion.
Left to right is time
Three minutes of music squeezed into one image: each column is a moment of the song.
Bottom to top is pitch
Bass hugs the bottom edge. Cymbals and breathy sounds rise toward the top.
Brighter is louder
A dark patch is near silence, and the brightest strips are what you hear most.
Eight seconds, up close
About as much as this demo can make. Two moments from our generator’s songs, at the zoom a Riffusion result comes in.
Columns are hits
Every tall stripe in “Forward Motion” is a drum. The evenness of their spacing is what makes it feel steady.
Riffusion today
What we found on 5 November 2026.
- Made by
- Seth Forsgren and Hayk Martiros, 2022
- Original code
- No longer maintained
- Original demo
- Showing an error when we checked
- This demo
- A community copy: prompt, negative prompt, 5 to 10 seconds
- Output
- A spectrogram image and its sound
- Paid use
- Allowed under CreativeML OpenRAIL-M, with limits on how it’s used
Riffusion questions
Is Riffusion the same as the Riffusion website?
No. The open model here dates from 2022. The riffusion.com address now redirects to a site called Producer.ai, and what that service runs isn’t what this model’s page describes.
Can Riffusion make a full song with vocals?
Not from what we could confirm. The model’s page describes turning text into spectrograms, and this demo stops at 10 seconds. Any claim of full songs is about something else.
Why would I still try a 2022 model?
To see the idea at work. Watching a prompt become a picture and then a sound is one of the clearest ways to learn how music AI began, and it’s free.
Who runs this Riffusion demo?
A Hugging Face user called fffiloni. It isn’t an official one, and neither its makers nor our site can promise it’ll stay up.
Not covered here.Ask us something else
Want minutes instead of seconds?
Describe a song and our maker writes the whole thing, from first note to last.
Make a songTwo free songs, no account needed.