Skip to main content
Read page text
page 38
or other text-to-image competitors, Stable Diffusion is an open source project, so its code is available for anyone to use or adapt. “The great thing with Stable Diffusion is that the guy who set it up, Emad Mostaque, is very ideologically driven in terms of the accessibility of artificial intelligence systems, and the importance of interrogating biases within the data sets. So a core part of the ethos of the company behind Stable Diffusion is that they wanted it to be open source, meaning that individual users could take this framework, and then apply it to whatever they wanted, and change it, develop it,” Roach explains, “Within weeks, so many innovations were taking place. People are making all sorts of different APIs, different applications that could use this very, very powerful system, and it’s been wild since then. Every other week you’re seeing things, and thinking, well, that was the stuff of dreams, and this is now possible. “What Stable Diffusion did has really changed the culture and the ideology behind what it means to have or to not have access to these tools, and radically opened up who can use them.” In late December 2022, shortly after Riffusion launched, Roach became fascinated by its possibilities – he had no plans to make an album, but found the idea of creating music with AI too compelling not to. “I started experimenting with it, and realised that there was a lot that could be done,” he says. “I got obsessed with this thing, and made hours and hours of recordings, because it’s quite lo-fi as well, it’s just this website that would spit the stuff out. I was documenting it – this is another thing I’ve been doing sometimes, because these systems, you don’t necessarily own them and you don’t know if it’s going to be there tomorrow. So I was making all this stuff and recording it, to return to it later, already with that knowledge that it’d be about going back, editing and finding things. I suppose the other thing too is that the majority of what was being spat out by the system wasn’t what I was looking for, even with the prompts I would use, it would take a long time to find something that really pulled me into it.” What is that something, in those snippets of generated sound, that made them just right? Were there specific prompts that worked, or failed? He demurs, and fair enough. Maybe the potential for text-generated audio using AI is greater than for text-generated images, because there is no standard of photorealism to compare the results to. Or maybe language is slippery and can’t be contained by the Venn diagram mentioned earlier. Music hits differently. Like the crate diggers zeroing in on treasure, you just know when it’s right. Or as Roach says, “I was talking with a friend recently about the pieces of music that were the most touching. Both of us were thinking about this. And I think there’s music that you can understand why you like it and you could talk about it forever. “But I feel like the most touching things, there are no words to say why, it just does something to you. It just does something to you, you know? And you know there are these attributes you could point to and other pieces of music that are structurally and sonically similar, or where they sit in a historical continuum. But then for some reason, this part of this song, the way the textures and the melodies or whatever it is work together, it just hits you in a way that is pre-linguistic and it touches you as a squidgy sentient person for some reason. “I feel like the [Riffusion] system can make things almost like pointing. Like, this is a free jazz drum solo, and this is an arpeggiated synthesizer, and this is a tabla. This is Rihanna. You can The Wire / patten 38
page 39
point to all of those things. But then there’s an elusive something else that isn’t there… as someone who makes music sometimes, it was that search for something. You’re playing on the keyboard, or you’re playing guitar or whatever you’re doing, you’re working with patches and looking for something to happen, you know. It’s always about that search to some extent.” After a short period of choosing text prompts and generating samples, Roach sifted and edited hours of AI-generated material, searching for bits with that elusive it-factor, before embarking on an almost forensic compositional process, deconstructing them, adding filters, delays, fades, EQ and pitch shifts, and patching everything together. The music is aided by the magic of the text box, yes, but the end result is very much one of personal sensitivity, taste and choice. “You could give someone else the same material and they would make a totally different record,” Roach agrees. “I’d like to show people what it is that they’re listening to, because it’s hard to know, isn’t it, what’s intentional. There are lots of things in there, like there isn’t a click or a glitch on that record that isn’t meant to be there. I wanted to talk about a surface that wasn’t perfect, where you can see the joints. Let’s talk about the idea of a thing, an object as an assemblage of parts or something. And also the idea of engaging with music in a different way, thinking about what the musical experience can be. “Also, in practical terms, there were a lot of limitations, in terms of how to make these into songs,” he continues. “I think one of the easiest pieces to talk about in that way is a track called “Drivetime”. Probably one of the straightest pieces on there. It’s a techno thing… but a lot of that has been made by taking the same bit of sound and then layering it and isolating tiny bits of the frequency spectrum to give that illusion… there’s a track called “In Me” where I’m juggling a smallish, shortish loop that I’ve created and then playing around with it like it’s plasticine or something, which I love to do, but I also tried to not do that too much on the record.” Elsewhere, in “Don’t Worry”, he smears a disturbed, distorted horn sound into a vocal that is just about singing the words of the title. The grime cadences of the vocal in “Walk With U” are sharp and staccato, set against jolting icebergs of flutey synths – keeping a steady flow, but devoid of intelligible lyrics. There’s a lot going on here, but it’s hard to say what proportion of that activity is happening within the studio or inside Riffusion’s models. “I did try to not get too technical about it because – how do you define too technical, you know. I did massively go in on it, but I wanted everything to feel appropriate. I wanted to use the lightest touch I could, and maybe also to kind of serve the material and try not to get in the way of allowing that material to live. “So in that sense, there’s an analogy with a sort of inverse archaeology, where you’re the custodian of this data. And you’re finding a way to deliver it, in both the sense of a postal delivery, but also in a maternal way, you’re trying to find a way to deliver this thing. It’s not you, it’s not about you. So that’s always been a big part of the patten project as a whole, trying to sidestep that centering of the self in making things.” Roach marvels at how far visual generative systems have come in the last nine months: “I’ve been surprised with the fact that you essentially have a text box that is inviting you to go and dream. It’s like well, what do you want to see? It’s anything you can imagine. Ask me and I’ll produce it for you. And what’s interesting is that what you actually tend to see out there are a lot of tropes from cinema and very familiar popular culture fused in some confident, combinative way. So it’s like Indiana Jones on Mars or Star Wars in the 1920s. It’s really strange. I think what’s really interesting to take away from that is to look at what forms our cultural imaginary... people have been given that opportunity to dream and think of whatever they can, and those dreams are often caught up in a very specific conceptual space, which is something to think about. “There are lots of questions about the potential of using artificial intelligence as a tool within the creative act in some way. And the possibility of it being easier than ever to make compelling images with these systems. What I am really intrigued by is the idea of – this is slightly different to the audio thing. But the idea of taking away some of the barriers to entry in terms of skill, even the financial privilege of time, to be able to sit and make something. The idea of taking all of that away as much as possible, where it’s really just a person and their ideas, and that’s where the value lies. These tools can be very directly used to interface with culture, not just as a consumer but as a producer of things. “I know there are lots of complicated ethical implications when it comes to the impact that some of these systems might have. Also obviously I don’t like the idea of giving our ability to make things, or to think about things, or to engage with each other and culture as a whole, to companies that don’t necessarily have the same ethics or any ethics. And it relates to Instagram, to Twitter, to the internet as a whole. That’s an existing conversation that then continues – we can’t pretend that it’s all bright skies and utopia with AI.” Another thing to consider with generative AI is how much the weight of history comes in. Will it become more difficult to produce truly new sounds if AI tools can only generate responses based on material in their training data set? While there’s a lot to be said for plunderphonics and sample based composition, surely the novelty of an inundation of whatever the sonic equivalent of Indiana Jones On Mars is will wear off fairly quickly. But AI can be used to piggyback on recognisable sonic tics, as it outputs bits of the cultural language of music that are ubiquitous, agreed on in their labelling terminology. Roach mentioned Rihanna, and there are pop diva vocals that emerge on some tracks with recognisable timbres and styles. Even if “You’re finding a way to deliver the data, in both the sense of a postal delivery, but also in a maternal way. It’s not you, it’s not about you” there aren’t words that are intelligible, they feel familiar, but also uncanny. “I don’t think this is like a dance music record,” he says, “There’s something exciting about musical formulas, in that certain musical formulas are so powerful that they will allow you to like you say piggyback, or use them as a trojan horse, for all kinds of weird sonic stuff. Like hiphop or techno or whatever are really good for that. Like really powerful carrier signals. A ballad, even. If you listen closely to a Drake song, there’s some pretty weird stuff going on, buried in the background, you know?” There’s also a human dimension to Mirage FM that continually makes itself felt, beginning with the glossy, almost vaporwave fantasy image of a red 1980s sports car enveloped in a swirl of exhaust on the cover (itself produced with the assistance of AI). From the emphasis on the idea of drivetime, and the sense of propulsion, tuning in to a radio playing familiar and ephemeral pop songs, to the feeling of being there, alive in a body, cruising through time and space and memory, there’s a tension between that form of experience and the disembodied data collection and connection of the AI. “It’s really important to bring up this thing of siting the whole album within the logic of the idea of radio, and also the listening that takes place in a car or while you’re doing something else,” says Roach. “This cloud based way of being, it’s like osmosis. I think what you could say is that the information, the artefacts like music, words and images that people throw into the stream of stuff we call culture, are traces of those behaviours. It’s like what's there in the space between all of the things in the museums and in books, in films and in the record racks. And like radio waves, still travelling through space, forever... in the space between all those things is a sketching out of our lives, of what it is that we do, what we dream, and what we feel.” In that latent space there is infinite humanity to be excavated, infinite crate digging to be done. ○ patten’s Mirage FM is released by 555-5555 39 The Wire / patten

Page Text

View plaintext

or other text-to-image competitors, Stable Diffusion is an open source project, so its code is available for anyone to use or adapt. “The great thing with Stable Diffusion is that the guy who set it up, Emad Mostaque, is very ideologically driven in terms of the accessibility of artificial intelligence systems, and the importance of interrogating biases within the data sets. So a core part of the ethos of the company behind Stable Diffusion is that they wanted it to be open source, meaning that individual users could take this framework, and then apply it to whatever they wanted, and change it, develop it,” Roach explains, “Within weeks, so many innovations were taking place. People are making all sorts of different APIs, different applications that could use this very, very powerful system, and it’s been wild since then. Every other week you’re seeing things, and thinking, well, that was the stuff of dreams, and this is now possible.

“What Stable Diffusion did has really changed the culture and the ideology behind what it means to have or to not have access to these tools, and radically opened up who can use them.”

In late December 2022, shortly after Riffusion launched, Roach became fascinated by its possibilities – he had no plans to make an album, but found the idea of creating music with AI too compelling not to. “I started experimenting with it, and realised that there was a lot that could be done,” he says. “I got obsessed with this thing, and made hours and hours of recordings, because it’s quite lo-fi as well, it’s just this website that would spit the stuff out. I was documenting it – this is another thing I’ve been doing sometimes, because these systems, you don’t necessarily own them and you don’t know if it’s going to be there tomorrow. So I was making all this stuff and recording it, to return to it later, already with that knowledge that it’d be about going back, editing and finding things. I suppose the other thing too is that the majority of what was being spat out by the system wasn’t what I was looking for, even with the prompts I would use, it would take a long time to find something that really pulled me into it.” What is that something, in those snippets of generated sound, that made them just right? Were there specific prompts that worked, or failed? He demurs, and fair enough. Maybe the potential for text-generated audio using AI is greater than for text-generated images, because there is no standard of photorealism to compare the results to. Or maybe language is slippery and can’t be contained by the Venn diagram mentioned earlier. Music hits differently. Like the crate diggers zeroing in on treasure, you just know when it’s right. Or as Roach says, “I was talking with a friend recently about the pieces of music that were the most touching. Both of us were thinking about this. And I think there’s music that you can understand why you like it and you could talk about it forever.

“But I feel like the most touching things, there are no words to say why, it just does something to you. It just does something to you, you know? And you know there are these attributes you could point to and other pieces of music that are structurally and sonically similar, or where they sit in a historical continuum. But then for some reason, this part of this song, the way the textures and the melodies or whatever it is work together, it just hits you in a way that is pre-linguistic and it touches you as a squidgy sentient person for some reason.

“I feel like the [Riffusion] system can make things almost like pointing. Like, this is a free jazz drum solo, and this is an arpeggiated synthesizer, and this is a tabla. This is Rihanna. You can

The Wire / patten

38

read next page

My Bookmarks

Buy this Book
Skip to main content