PSA: THE LINK TO THE RIFFUSION WEBSITE HOSTS A COMPLETELY DIFFERENT CLOSED MODEL - MISLEADING
Figured its a good idea to let people know this since the author isnt fixing the issue.
The experience at https://www.riffusion.com/ is NOT Riffusion-model-v1, but a completely different generator that we do not have access to, same as Suno and the others.
The riffusion.com website is just yet another one of those groups who steal copyrighted content from real musicians without license or permission (we the producers often use frequency signatures in order to ensure any AI will end up also replicating that, so that we can find out easily if an AI is trained on our music without permission), in order to profit off of it with online subscriptions and misleading information. Using riffusion.com gets you generations that are based on UNLICENSED music, and trying to distribute something like that might get you sued.
The riffusion model hosted on this Huggingface repo isnt all that bad, compared to things like Stable Audio Open V1.0 for example, but it is lightyears away from what you might expect after generating something through their website.
To Riffusion: If this discussion thread is removed or made inaccessible/hidden without the issue first being resolved, alternative methods may be enacted to ensure the problem is fixed and/or the public is aware of the issue.
Best regards,
Chris | LiVeen + A community of ~200 producers, A&R, labels, distributors and vocalists
I'm confused. What is it exactly that you are asking/demanding here?
Because from my perspective, you are talking about many different things at same time.
Some of those things may be valid concerns, if true (such as weather unlicensed commercial music was used to train their web model). However, it's easy to jump to conclusions without doing a proper analysis. I have actually tried to get the model to reveal its training secrets, and so far not successful. Whereas this has been quite easy with some of their competitors out there, where I can easily get the models to reproduce commercial music with original vibes and melodies and rhythms. That said, it would've been really nice if when making a public statement like this, you also provided some specific examples. Without that, I see no reason why anyone reading this post wouldn't just dismiss it?
Some others, however, seem to be a specifically a you-type "problem" (such as your accusation/ complaint that this Huggingface repo over here is somehow misleading, whereas no person familiar with Riffusion would think that this model, available for download in here, while being 2 years old, is somehow being used in the backend to generate high quality music on their website (which, by the way, only recently opened up to the whole world) 🤷♂️
So let me ask you again, what exactly are you demanding?
I personally think that you're not actually a serious person, otherwise you would have contacted them with your concerns directly, instead of creating this mess of a post in here.
"
Riffusion
Riffusion is an app for real-time music generation with stable diffusion.
Read about it at https://www.riffusion.com/about and try it at https://www.riffusion.com/.
"
This is copy pasted from the very top of this repo.
When I made my comment, this was further down. It's even worse now, and apparently you think that this DOESN'T imply that this repo's model is what's used on the riffusion.com website?
Seriously?
I want them to stop acting like the riffusion.com website's model has anything to do with this one. It doesn't.
As for unlicensed music in the dataset of the model hosted at Riffusion.com - They'd have an opportunity to make a hell of a lot more money if that one was made legally. Instead, they're forced to do things the same way as Suno, because access to the model file would lead to people finding straight-forward, conclusive evidence of unlicensed music. Instead, they don't even try to claim that it's legal on their website.
Hell, even this repo's "dataset" section is phrased in such a way that they're free to argue they never claimed any of the linked sources were actual datasets used, which just seems weird to me.
As for contacting them with our concerns directly: We did. They don't care, and don't even listen. That's why we're here.
The part I care most about here is the extremely misleading model card of this repository, and it's so damn easy to avoid, and yet they do not care.
The fact that you're trying to defend them like this is very concerning to me.
Best regards,
Chris | LiVeen
I’m doing some research to try to find clues about how the Riffusion models were structured and how they actually work — mainly Fuzz 0.8 and 1.0. Today we also have 1.0 Pro, 1.1, and 1.1 Pro, which are no longer the same thing — they improved in audio quality but the generations became more inconsistent.
As for the sample model in this repo — it’s definitely nothing close to the older models they used to have. It’s more like a prototype or a “placebo” version of the original idea — something that in practice evolved into something else entirely (at least that’s what it looks like).
Regarding how the model is trained, I honestly think all this debate about legality is a bit silly. It’s like saying a musician shouldn’t listen to music in order to learn how to play and reproduce a style or genre. As long as you’re not making exact copies or committing plagiarism, it should be fine. The same logic applies to models — they need references to learn how to replicate something according to the learned patterns.