It's finally out
Cinderella After the Spell Broke is finally out. It's the music video that took the longest and gave me the most trouble so far.
Lyrics that started with a friend
This song started a few years ago, when an old friend of mine handed me a rough draft of the lyrics. I let it sit for a while, then started working on it bit by bit about two years ago.
It was different from lyrics I write on my own. The words and phrasing didn't fit my usual mold. I tweaked a few spots, but looking back now, I think I managed to turn it into a song almost as is, without breaking the image behind it.
Compromise is hard with AI
Given that history, making the music video meant doing what sounds like the ordinary job of turning a song into visuals. But with AI-generated images and video, I ran straight into how hard it is to compromise. In the past, getting about 80% of what I pictured was good enough. This time, the moment something felt even slightly off, I had to redo it, over and over.
The one-ear wired earphone problem
AI really does struggle to lock down small details. The wired earphone worn in one ear, which is central to this piece, was especially badly handled. A single generation always came out with contradictions or something that felt off.
The same goes for the shape of hands and fingers. For details like these, a two-step approach works overwhelmingly better: generate the rough composition first, then fix the details afterward.
For example, first I output the person without an earphone. Then, starting from the ear, I erase the path the earphone cord would take, as if with an eraser, and give instructions for just that area. That way the composition and the person, which the AI isn't that weak at, stay fixed, and I can regenerate only the earphone as many times as it takes until I'm happy with it.
Making people in their school days, 20s, and 30s
Another thing: this time, the same two main characters appear from the first frame to the last. And I needed three versions of them: school days, their 20s, and their 30s.
I suspect that foreign-made AI doesn't really understand the ages of Japanese people. A 30-year-old foreign person and a 30-year-old Japanese person look quite different. So I tried prompts that lowered the age, and when I was close I'd fix just the "wrinkles." Even producing the four-view character sheet took trial and error.
Arrangement and Suno's noise
I adjusted the arrangement in SunoAI again this time. But when a song has distorted guitar like Noise Margin's, noise creeps in during the second half and the vocals start to distort, as usual. I've tried a lot of things, and for now I'm getting by with tools like iZotope RX 8's De-clip.
I still feel Suno's stems (the separate parts of the track) aren't accurate enough. I just hope future feature updates don't go off in some strange direction. What I really wish is that it generated from stems in the first place, then mixed them for output. Still, I have no sense for arranging, so in the end I'm leaning entirely on Suno.
What's next
As for what's next, once I catch my breath I plan to dig into my YouTube analytics. My impressions are weak, and YouTube clearly doesn't understand my channel yet. I want to show it, loud and clear, that this is the kind of song I keep putting out.
I'll keep going steadily, for myself, with no regrets, so thank you for your support.