Talk:Riffusion

Alternate audio example

I added the "bossa nova with electric guitar" example to the article because the width of the spectrogram image is smaller (and therefore easier to fit in the article while still being relatively readable), but in case there's a preference for something else for whatever reason, there is another alternate example:

Generated spectrogram from the prompt "1980s synthwave pop" (top), and the resulting audio after conversion (bottom)

This one has longer audio duration, but that also means that the spectrogram dimensions might be awkward for the article. I'll leave it to the rest of you guys, though. --benlisquare_T•C•E 03:41, 18 December 2022 (UTC)

Updates

Hi. My name is Kendall and I work for Riffusion. In compliance with WP:COI, I wanted to request an impartial editor consider the below proposed changes:

−

Riffusion is classified within a subset of AI text-to-music generators. In December 2022, Mubert similarly used Stable Diffusion to turn descriptive text into music loops. In January 2023, Google published a paper on their own text-to-music generator called MusicLM.

+

Originally Riffusion was a hobby project created in Forsgen's basement. They formed a startup, also called Riffusion, and raised $4 million in venture capital funding in October 2023. They released an upgraded version of Riffusion with a new AI model that can create entire songs, rather than just short phrases. In 2025, Riffusion released an AI model called Fuzz, which creates music from a variety of text, audio, or visual prompts, and considers the user's musical preferences. It has settings for instrumentation and style. As of January 2025, Riffusion is still free-to-use.

New Content with Citations In-Line

Originally Riffusion was a hobby project created in Forsgen's basement. They formed a startup, also called Riffusion, and raised $4 million in venture capital funding in October 2023. They released an upgraded version of Riffusion with a new AI model that can create entire songs, rather than just short phrases. In 2025, Riffusion released an AI model called Fuzz, which creates music from a variety of text, audio, or visual prompts, and considers the user's musical preferences. It has settings for instrumentation and style. As of January 2025, Riffusion is still free-to-use.

References

References

Explanation: The citations in the last paragraph of the current page do not seem to even mention Riffusion (maybe I missed it?). The proposed replacement paragraph would help keep the page up-to-date with new sources about the Riffusion business and version 2 of the software.

Kendallrankin (talk) 18:02, 27 February 2025 (UTC)

Partly done: I took out the promotional parts, but added. Likeanechointheforest (talk) 21:21, 28 February 2025 (UTC)

Thanks @Likeanechointheforest:. Some of the proposed additions were intended to add corrections resulting from out-dated content about version 1 of Riffusion by adding information about version 2. Maybe there's a better way to address the errors/outdated info? Here are the issues:

1. "This results in a model which uses text prompts to generate image files" -> The new version can use text, audio, or visual prompts, not just text prompts.
2. "only several seconds long" -> The new version can create entire songs, rather just a few seconds.
3. "It was created as a fine-tuning of Stable Diffusion, an existing open-source model for generating images from text prompts, on spectrograms." -> Riffusion no longer uses Stable Diffusion, spectograms, etc. and can now use visual and audio prompts too. There aren't really press articles saying Riffusion no longer uses these methods, just articles saying a new model was introduced.

Basically, the issue I'm trying to resolve is that the entire page is about version 1 but uses present tense and doesn't explain this is historical information not accurate for version 2 (the current version). Kendallrankin (talk) 17:26, 20 March 2025 (UTC)

Thank you for pointing this out! Typically, granular details about software or tools are not deemed encyclopedic, so I've taken it out rather than necessitate constant updates. Likeanechointheforest (talk) 22:26, 21 March 2025 (UTC)

I have restored the stuff you took out; I don't really understand your reasoning here. I am quite familiar with the policies you're gesturing at, but I do not know of anything resembling a general prohibition on describing how software works or what it does. If there is a new version of the software, then we can describe what it does and how it works as well: this does 'necessitate constant updates', but that is generally what happens when you write articles to contain useful information rather than generic PR stuff.

@Kendallrankin: I have clarified that the original version has been superseded; is there any actual information on this "Fuzz" that replaced it? Is the source available, or is there at least something resembling a whitepaper? jp×g🗯️ 05:42, 26 March 2025 (UTC)

@JPxG: Journalists have written about the new version (Fuzz). Below is the content I previously shared about the new version, including citations.

New Version Draft Content

They released an upgraded version of Riffusion with a new AI model that can create entire songs, rather than just short phrases. In 2025, Riffusion released an AI model called Fuzz, which creates music from a variety of text, audio, or visual prompts, and considers the user's musical preferences. It has settings for instrumentation and style. As of January 2025, Riffusion is still free-to-use.

References

Is that what you were looking for? Naturally, I'll leave decisions to impartial editors, but happy to help however I can.Kendallrankin (talk) 17:03, 3 April 2025 (UTC)