zzzzzzzzzzzzzzzzz
4 Sept 2025
#process #motion #AI
This post feels like a Simpsons episode: the beginning has nothing to do with the ending. It all started as an exercise in a creative AI course. The dynamic was to play with the srefs in Midjourney (numbers that define the style of the images). With the sref random command you get random styles, and if you like one, you save it. Easy in theory, but then you get carried away generating images… and I ended up making 111 girls blowing a bubblegum bubble. Yes, one hundred and eleven. Because none of them felt quite right.
At that moment I thought I had wasted hours, that I was too picky and obsessive, that I needed to be more practical and just finish things or I won’t make a living at this… in short, I had a pile of images that were useless. But later (plot twist), that “useless material” turned out to be exactly the piece I needed.
What seemed useless ended up being the key
Meanwhile, on LinkedIn…
Aaay LinkedIn. I don’t log in often, and when I do I try to focus on the few people actually sharing useful, practical stuff. But it’s not easy… between self-proclaimed AI gurus, empty quotes, humble brags and endless office moans, the noise is overwhelming. And hey, that’s fine, everyone can use their socials however they want. But if your feed is just recycled “visionary” takes and grand declarations, it makes you not even want to read (or listen). And since life is full of contradictions, here I am dropping my own big, wannabe visionary statement, in blog highlight form:
By the time you finally have something valuable to share on LinkedIn, nobody’s listening, because you’ve already bored the algorithm to death
From there (and a bit of hate towards social media, let’s be honest), came the connection: everyone saying the same thing on LinkedIn, inflating the bubble (myself included), until it pops. And when it pops, nothing is the same. And then I remembered my 111 bubblegum girls. What seemed like a waste of time ended up being the main visual piece behind POP ✨.
I had it, but I was still missing THE scene: the bubble popping. And here comes the part people don’t usually tell about AI. They only show the pretty results and it looks like magic. Well, either I’m really clumsy (and I even used a custom GPT, the AI Video Prompt Generator for Veo2, Kling, Runway, to refine the prompts) or it’s not magic: it’s trial and error, spending credits, and crossing fingers that the model understands something from your prompts. I thought: “pfft, a bubblegum bubble popping, that must be easy for AI.” Nope. What I got in many tests was even disturbing (don’t watch them before bed).
AI is iteration, it’s trial, it's error and lots of outtakes
I ended up splitting the action in two: first the girl blowing the bubble (the most convincing result came from Kling 1.6 standard) and then, with VEO2 and plenty of patience, the explosion and splashes.
I learned there’s no single model that works for everything: the best one is the one that adapts to what you need right then. Next time, I promise to do better research on which model does what before I waste attempts like a maniac
Once I nailed the bubble, I moved on to the script and audio. I worked the script with ChatGPT to shape the narration. With version 4—not the standard one, the more creative brainstorming-friendly one (don’t ask me which exactly, because it no longer exists). For audio I used ElevenLabs with the voice Ivanna – Girl Next Door, which sounds very natural, not like an ad or a podcast host. Though again, it took some tests, because you don’t get the voice you want on the first try.
I added a café background murmur (free download) and random AI phrases with other ElevenLabs voices. Result: a messy soundscape and a casual, realistic tone.
No matter the AI hype, you still need the classics.
Then came the classics: because in the end, the old tools are still the ones that give you full control over the result and the final polish. I used After Effects timewarp for the acceleration in the sequence of girls, over a duplicated layer, a circular mask selecting the bubble and scaling it to simulate its growing, I also added a wiggle to mimic the imperfect motion you get when blowing a bubble. I blurred the center of the bubble too, since some mouths looked way too sharp and to big after scaling to be believable.
And Premiere for the final edit, subtitles (yellow on black, always, for readability), and audio mix with Essential Sound to give everything unity and polish.
And that’s how POP was born, a 25-second micro-story that looks simple but hides a ton of tests, errors, tweaks and learnings.
What I take away:
Many ideas that seem creative actually come from paying attention. “Chance” isn’t chance: it’s exercising those absurd connections between things that at first seem unrelated.
AI doesn’t get you on the first try: it’s more like a process of iteration, finicky and shifting. You have to stay flexible and learn to salvage whatever is useful from what it throws at you.
And personally, not to be so anxious or beat myself up over what I think I did wrong. Trust the process: what looked like wasted time turned out to be exactly what I needed.
And now that you know everything about the process, here's the final result. Enjoy it :).
Read Next
