11 votes

Riffusion – Stable Diffusion fine-tuned to generate music

6 comments

  1. [4]
    stu2b50
    Link
    It’s not bad, considering it’s a much more hobbyist endeavor than the other generative nn projects. They basically just trained stable diffusion on spectrograms. With more resources, both human...

    It’s not bad, considering it’s a much more hobbyist endeavor than the other generative nn projects. They basically just trained stable diffusion on spectrograms. With more resources, both human and compute, on the level of what gpt3 or the various art generators get, it’ll be much better now that people have a proof of concept.

    Pretty fun to play with anyhow. At this point you could probably make a full indie game just from “AI” - art assets from stable diffusion (probably would want pix2pix so you can control things somewhat), music from this, plot from gpt3, programming assistance from copilot.

    7 votes
    1. [3]
      Omnicrola
      Link Parent
      It would be a really interesting project, though I fear the backlash that such an undertaking (no matter how well-intentioned) would receive at this point. Rather than making a game, I think it...

      At this point you could probably make a full indie game just from “AI”

      It would be a really interesting project, though I fear the backlash that such an undertaking (no matter how well-intentioned) would receive at this point.

      Rather than making a game, I think it would be more interesting to try and come up with an "intelligent" immersive experience. So while the overall direction, theme, and tone could be set by the author, the actual output could be dynamically tailored in real-time to adapt to the person experiencing it. ..... shit did I just describe a holodeck?

      1 vote
      1. [2]
        stu2b50
        Link Parent
        Yeah and I think that's a shame. I was in NY MOMA recently and a lot of the impressionist art was based around "removing" the artist from the process - things like paint splatter. Randomness as a...

        Yeah and I think that's a shame. I was in NY MOMA recently and a lot of the impressionist art was based around "removing" the artist from the process - things like paint splatter. Randomness as a way an interesting way to play with the very concept of having an artist.

        So it's a bit of a shame that a "full AI" game as just an art experiment like those would have backlash for what seems mostly tribalistic at this point.

        3 votes
        1. moocow1452
          Link Parent
          AI Dungeon can already make generative art, rope in some automatically generated music, I think you could do it pretty easily.

          AI Dungeon can already make generative art, rope in some automatically generated music, I think you could do it pretty easily.

          2 votes
  2. Wulfsta
    Link
    Wow, this is a lot of fun to play with.

    Wow, this is a lot of fun to play with.

    2 votes
  3. Akir
    Link
    It's weird; I think the fact that the training data was compressed lead to the algorythm occasionally producing noisy outputs. It's also really weird when you enter a famous singer because you get...

    It's weird; I think the fact that the training data was compressed lead to the algorythm occasionally producing noisy outputs. It's also really weird when you enter a famous singer because you get their voice and the vocalizations that they are famous for.