How to get AI to think outside of your box?

[BBFG#]

Seems to me the first thing to endeavor with these are the prompts themselves. After watching some documentaries on this, I noticed for music that more than a few start with things like genre, style, mood, tempo etc. as the "primaries". And then things like loops and phrases in rhythm, melodies, etc.
It all struck me as building a box to compose in without the builder realizing they can only build a box from their limited knowledge of boxes.
First, how do we get something going without pigeon-holing it into a genre, style or mood? And isn't mood a subjective term anyway? I know the movie subtitles we watch make us laugh when it dictates what the music mood is.

How can we prompt AI without building a limiting parameter box?

[ghettosynth]

Seems to me the first thing to endeavor with these are the prompts themselves. After watching some documentaries on this, I noticed for music that more than a few start with things like genre, style, mood, tempo etc. as the "primaries". And then things like loops and phrases in rhythm, melodies, etc.
It all struck me as building a box to compose in without the builder realizing they can only build a box from their limited knowledge of boxes.
First, how do we get something going without pigeon-holing it into a genre, style or mood? And isn't mood a subjective term anyway? I know the movie subtitles we watch make us laugh when it dictates what the music mood is.

How can we prompt AI without building a limiting parameter box?

Break the genre down. There's an example in the video that I posted relating to the discussion of 8-bit music where I showed the different results for "chiptunes" from just asking for the genre and being more specific.

This is the video. Keep in mind that this is google's generator, and I'm not particularly impressed by these results, but, the prompts are in the youtube description.



The tradeoff that you will run into is between more constraints giving more focus, but, more constraints constraining the model in unexpected ways. It's trying to find a plausible path, it is not a constraint solver. Hence, the more detailed that you get, sometimes, the less it is able to obey your wishes.

So, if you have more skill in specifying music constraints succinctly, you can get the model to produce music closer to your intent. Also, the thread that I started in this forum about what other people like is intended to explore this. That is why I posted the prompt that was used so that there would examples that describe the music, but that you can hear the music, and that the music is moderated by collective taste.

[BBFG#]

What if l you're exploring an unknown genre? Most genres we have today were created after the fact and not before.

But thanks for that video. Always up to get more perspective on this.

[BONES]

First, how do we get something going without pigeon-holing it into a genre, style or mood?

The simplest solution is to reduce the adherence to your prompt. SUNO's default is just 25%, which gives the AI plenty of scope to go outside of the box you create for it.

[BBFG#]

Even that feels restrictive though.
I realize that many "genres" are quick easy dumping grounds for the markets they hope will consume them. So maybe play the ideas first and then not allow AI coin it under those easy umbrella terms?
Can we frustrate it into evolving?
(Instead of simply "learning")

[ghettosynth]

Even that feels restrictive though.
I realize that many "genres" are quick easy dumping grounds for the markets they hope will consume them. So maybe play the ideas first and then not allow AI coin it under those easy umbrella terms?
Can we frustrate it into evolving?
(Instead of simply "learning")

You don't have to give it a genre, per se. Just use language to describe the music that you want at whatever level of details that matters to you and that is reasonable within the prompt length limits. Just don't expect it to treat your prompts as a set of constraints.

[Kenmac]

I've posted some prompts here:

viewtopic.php?t=628223

Feel free to try them or share your own if you want to.

[BONES]

Even that feels restrictive though.
I realize that many "genres" are quick easy dumping grounds for the markets they hope will consume them. So maybe play the ideas first and then not allow AI coin it under those easy umbrella terms?
Can we frustrate it into evolving?
(Instead of simply "learning")

The thing with the way it works, though, is that you get the best results when you are very specific in what you want it to do for you. Without that specificity it is just going to go the route of the most popular stuff. I don't think it is set up to think outside the box because I think that would frustrate a lot of users who just want something to listen to. You have to remember that it is not a tool made for musicians or song writers, it's made so that absolutely anyone can make their own music, which means skewing to familiar things.

[Michael L]

you get the best results when you are very specific in what you want it to do for you.

Can you share a sequence of your very specific prompts, to show how you get Tunee to give results that express your own musical ideas?
I cannot comprehend how one would compose music using descriptions.

[BONES]

I don't do the AI stuff, that's all my bandmate. I'll see what he's willing to share but you have to realise that it's not just his prompts, he keeps at it until it spits out something useful. His hit rate is probably less than one useful thing in a hundred. He's generated more than 100,000 images in Mid Journey, he's just a little bit obsessive about it. I don't have that kind of patience, I get bored after a couple of iterations and turn the telly on.

[Michael L]

His hit rate is probably less than one useful thing in a hundred. He's generated more than 100,000 images in Mid Journey, he's just a little bit obsessive about it.

So, Tunee outputs dopamine-inducing, variable rewards thereby making it hyper-addictive, like social media on 'roids.
I am so not surprised.
Thanks for sharing the IRL details.

[BertKoor]

This all sounds familiar, specific prompts don't work in other fields as well:

The Real Secret of Prompt Engineering
Since early 2023, I’ve been on a journey evaluating claims about the capabilities of generative “A.I.” (yep, still gets air quotes).

I’ve tried to reproduce some of the more sensational successes I’ve seen trumpeted on the Interwebs, and eventually come to the conclusion that most of them don’t hold much water.

Why, I wonder, are these people claiming to have done things that the technology just doesn’t seem able to do?

Their defence is typically that I must be “doing it wrong”; that I haven’t mastered the Secret Magical Prompts of Destiny. But when I try to follow the advice, I get the same “meh” results.

“Be more specific” is a common refrain. But here’s the thing:

• I’m a computer programmer with a degree in physics. I can do “specific”.
• If anything, the more specific the requirements, the more the models struggle. When I try to iterate the output through a longer conversation, the results can often get worse.

Over these two years, I’ve gradually developed a theory about how they’re succeeding with “A.I.” where I’m failing, and it’s probably best illustrated with a cartoon.

This image was generated through the ChatGPT web application (so it was generated by DALL-E, I guess). It went through multiple iterations as we tried to correct the problems, but – as often seems to be the case – the first attempt was about as good as it got.

I was very specific in my prompts about the story, the dialogue, the characters, down to the level of exactly what should be featured in each panel.

Some folks looked at this image and saw the continuity mistakes. Others looked at it and said “It looks okay to me”.

And that, I suspect, is the secret.

[eassae]

The title of this thread sound way dirtier than I think was intended?

[BONES]

Yeh, all that is definitely true but you have to be patient and learn how to get what you want from it. It's not straightforward, it's not easy but it's definitely doable.

To use a video example, just check this out. The detail is unbelievable and to create this using state of the art CGI pipelines like ILM or Weta Digital, you'd need a 100 artists working for several months. But the rata at which this stuff gets pumped out suggests it can be done by one person in a matter of days or weeks. It's unbelievable. Just look at this stuff -



Now, this is not to say that AI is ready to make feature films, it is still a long way from that but the thing is, this is how far it's come in what, two or three years - imagine where it will be in another two or three.

What I think this guy is doing is the same thing I think my bandmate does - neither are starting out with a specific aim, they just have a general idea of what they want. This guy wants to make interesting looking Future-Retro video scenes and Craig wants to make good EBM/Industrial music. In both cases they guide the AI - more of this, less of that - and let the AI take it where it may. Craig might tell it to use a faster or slower tempo, but he won't specify a bpm. Both of them are being creative on the fly, working with the AI to create incredible things, rather than enslaving the AI to a particular result that they have in mind.

Craig says there are lots of things it won't do, like use a particular key or make a track without vocals. Sometimes it might but mostly it won't so you work with it, you go with the flow. You nudge it into position, then you let it iterate and iterate until it does something worth keeping.

What I don't think you do is constantly rewrite a prompt when you don't get what you're after first go. If the first pass is in the general ball park, then try another dozen generations using the same prompt and see what you get. Remember, it's not science, it's art.

You can be specific with your prompts but you can't be too narrow with your goals. So, for example, Craig didn't set out to generate 12 particular songs for an album, he just set out to generate songs and selected things that worked, whether they were what he was thinking of when he started or not. That's how creativity works, you go with the flow and see where it takes you. Idiots like Zeisner, who want it to do an overly specific thing, will be disappointed but people who just want to make good music might find it very rewarding.

[Bunny_boy]

This all sounds familiar, specific prompts don't work in other fields as well:

The Real Secret of Prompt Engineering
Since early 2023, I’ve been on a journey evaluating claims about the capabilities of generative “A.I.” (yep, still gets air quotes).

I’ve tried to reproduce some of the more sensational successes I’ve seen trumpeted on the Interwebs, and eventually come to the conclusion that most of them don’t hold much water.

Why, I wonder, are these people claiming to have done things that the technology just doesn’t seem able to do?

Their defence is typically that I must be “doing it wrong”; that I haven’t mastered the Secret Magical Prompts of Destiny. But when I try to follow the advice, I get the same “meh” results.

“Be more specific” is a common refrain. But here’s the thing:

• I’m a computer programmer with a degree in physics. I can do “specific”.
• If anything, the more specific the requirements, the more the models struggle. When I try to iterate the output through a longer conversation, the results can often get worse.

Over these two years, I’ve gradually developed a theory about how they’re succeeding with “A.I.” where I’m failing, and it’s probably best illustrated with a cartoon.

This image was generated through the ChatGPT web application (so it was generated by DALL-E, I guess). It went through multiple iterations as we tried to correct the problems, but – as often seems to be the case – the first attempt was about as good as it got.

I was very specific in my prompts about the story, the dialogue, the characters, down to the level of exactly what should be featured in each panel.

Some folks looked at this image and saw the continuity mistakes. Others looked at it and said “It looks okay to me”.

And that, I suspect, is the secret.

I've just done that using the Google AI and added, "in the style of Francis Bacon" to the end. Looks nothing like Bacon's paintings more like an EC comic..!