Artificial intelligence

Normally, I don’t share too much information or detail on how I produce my art. However, when some art is made using generative AI tools, I think it is important to be totally transparent about how it is used to avoid any misunderstanding.

To be clear, in this explanation, I will only cover the use of generative AI and not other AI tools such as Stem Separation from Spleeter or ML Enhance from Pixelmator.

I think the reason generative AI is met with a fiery reaction by some people lies in the apparent simplicity and, more importantly, the speed of generating audio, images, videos, as well as code, 3D meshes, etc. It is clear that, from a production point of view, artists cannot compete against this machine by quantity, and as the quality of output increases, it will become even harder to compete in quality.

As of today (November 2024), I only use generative AI as a tool to generate content that I then modify, rework, and remix for my art. So far, only three pieces have used this technology for the Steo Le Panda art project:

For each of these pieces, the output generated by the AI tool was used as a building block to create something more evolved. It wasn’t just “enter the prompt, choose one I like, and post it unmodified as genuine work.”

I can assure you it wasn’t used out of laziness, as I spent a long, long time creating these artworks and remix albums (far more than I thought it would take, even with the use of efficient tools like these).

For example, in the artworks, multiple AI-generated images were fused together using traditional software (Pixelmator Pro) to create the final pieces. This is similar to a patchwork artwork using these generated images as resources, adding effects, and combining them. I could have obtained a similar result using stock photos (not AI-generated), but it wouldn’t fit the EP/album themes of cyberpunk in China.

For the remix album M4 CH1N3 QU1 R3V3, each track was condensed from between 4 and 12 generated outputs. Each of these outputs was created based on audio samples I fed the algorithm, sometimes just short extracts from one of my songs, sometimes just a few sounds.

However, most of the voices were added afterwards using original recordings, as the AI tool was not great at processing them without distortion. The only voices generated by AI for this work are on M0R3 C0W83225, 4D5, and 82155.

Generative AI for websites and apps

There is another area where I use generative AI a lot: building websites and apps.

Strangely, I feel much more comfortable using it there than for music. In music production, generative AI can sometimes leave me with the less fun part of the work: editing loops, arranging material, mixing, mastering, and trying to turn generated ideas into something that really feels alive.

For websites and apps, it is almost the opposite. When I was making websites with WordPress or Tumblr, I liked having a website, but I never really enjoyed building the pages. Even with a content manager, there were always layouts to fight, settings hidden inside other settings, and small technical problems between the idea and the result.

For apps, it was even worse. I always had many ideas, but trying to code them was often a nightmare. My panda brain is simply not naturally wired for traditional programming.

It is not that I hate logic or systems. I actually enjoy “code with noodles”: Quartz Composer, Max for Live, Unreal Engine Blueprints, and other visual tools where I can connect ideas together and see what happens. But writing regular code from an empty screen always felt much harder.

With generative AI, building websites and apps becomes much closer to what I enjoy: imagining the concept, designing the experience, testing the prototype, breaking it, improving it, and creating the art around it.

In this case, AI does not remove the fun part. It removes a lot of the boring wall between the idea and the prototype.

I think generative AI is an amazing tool that, right now, can help artists experiment with a lot of control and really push art creation in a good direction. However, it is true that it can also be used to generate soulless and lazy content. But in the same way, audio samplers can be used very creatively, or just in a lazy way (like sampling a famous disco song and only adding a house beat without any other production work…).

Additionally, I think in a few years/decades we will look back to these days with nostalgia, and the lo-fi sound quality of these early AI generative will be a stylistique effects that future producer might want to emulate, the same way they add “warmth” to songs by using cassette audio effects.

I do believe that being able to control the AI models—even being able to train them yourself—is the better way to go. This is an area in which I still have much to learn, but I am interested. For example, I would like to create a generative AI that could take the sound of my music and apply it automatically to the composition of another. Does this ring a bell?

(This article has been typed by hand on a computer, but sent to a LLM for checking spelling mistakes…)