Here's my short rant on this, if you want to just test the guide for yourself without reading, go ahead and do so, that is entirely possible. The guide .md file literally explains itself when dropped into your LLM of choice.
I started with a simple idea, take image, create song out of it. Now I have this entire framework from every empirically tested and best practices in prompting and music design for ACE-Step1.5 I could possibly create.
It is literally easier for you to just drop the guide into a LLM than for me to explain the entire thing, but here's the idea.
Drop the guide onto any LLM. You can now tell it your intent or ideas for music, and it will get to work. It can be pretty much anything you can think of. You could be like me and drop an art image. You could just describe the song you want. You could pick specific instruments you want. You could tell it to help you make a song mixing seven different media and five different hypothetical emotional or philosophical framings. I worked with LLMs to design this so that basically anything you could possibly think of, will produce SOMETHING. You can even tell it you want to edit the guide itself to fine-tune it for your own use cases.
Mechanically, it takes whatever your intent is, and translates that into its baseline values like music theory, references, emotions, whatever you gave it. It then turns that into examples for you to pick between and you can freely pick these or tell it to add things to them, include your own examples, mix them, whatever. It's an iterative design process with the LLM that will converge with whatever your desire is to eventually create a simple copy/paste Caption+Lyrics prompt into ACE-Step1.5.
I've tested this by literally cold dropping the guide with random images, and saying literally nothing, and it provides me with example prompts to build off.
Do note that if your intent is slow, methodical songs like slow jazz etc, I wish you good luck, because you're putting yourself on Dark Souls difficulty. Make Miyazaki proud.
If you have any feedback, it would be greatly appreciated, you can comment here because this is my personal production framework as well, so I'm always looking to improve it or fix any possible problems.
If you used this guide and ran into failure cases, I'd love to hear them, because it means you found an edge case that I'd want to try to fix.