zzzzzzzzzzzzzzzzz
#process #motion #AI
It all started with a discarded idea. Or rather, with 111 of them.
I was testing the --sref random command in Midjourney on an image of a girl with bubblegum, to explore styles and save the ones that might fit future projects.
In the end, I don't really know what happened, but I ended up generating 111 images of girls blowing a bubble. One hun-dred and e-le-ven.
Partly it was because I couldn't find what I was looking for, but I also admit I fell into the AI trap: that loop of doing just one more test to see if the next one comes out perfect. Since it was a personal project with no deadline, I could iterate to infinity. And so I did 😅
It felt like a waste of time and credits. Those 111 images sat in a folder for a while, gathering digital dust.
What seemed useless at first ended up being the key
One day, scrolling through LinkedIn, it clicked. There was so much noise, so many categorical statements, so much AI theory without any practical grounding, that constant sense of urgency not to be left behind... the tech bubble felt like it was about to burst. And that feeling matched my saved material perfectly. Those 111 useless images of different people blowing the same bubble became the perfect visual representation of POP's core idea.
I had the foundation. Now I "only" needed the scene of the bubble popping.
On social media, you usually see the clean, final result without errors (or with the flaws very well hidden). The truth is, asking the AI for an action as specific as "a bubblegum popping", I think I almost broke the model.
AI is all about iteration: trial, error, burning through credits, and crossing your fingers that the model actually understands your prompts
What physically seems like a simple action translates into a catalog of random deformations and glitches in video generation. And not for lack of technical precision: I used a custom GPT (AI Video Prompt Generator for Veo2, Kling, Runway) to structure the prompts. Even so, some results were straight-up unsettling. Here are a few with their respective prompts and models:
Ultimately, the solution was to stop chasing the "perfect prompt" and break the shot down by technical requirements:
Kling 1.6 Standard: I used this exclusively for the inflation action. It was the model that best understood the physics of the bubble expanding.
VEO 2: I used this for the moment of the burst and the gum splatters.
When people ask me which model is better for what, I always say there's no ultimate model. The winner is whatever model solves your shot at that exact moment.
With the bubble sorted, I shifted focus to the script and sound design. I shaped the narrative using ChatGPT (I used v4, the specific version that was actually good for creative brainstorming before they sunsetted it).
For the voiceover, I went with ElevenLabs, using the Ivanna , Girl Next Door voice. It feels organic, no "ad" tone or polished podcast inflections. Still, just like with the imagery, the voice rarely hits the mark on the first take, it took several trials and iterations to nail the tone.
To finish the soundscape, I layered in some background coffee shop chatter (from a free stock library) and mixed it with random AI-related phrases from other ElevenLabs voices. The result? An intentional, realistic sound clutter.
I feel like AI is incredible for starting things off. You no longer have to face a blank canvas, and it lets you generate ideas at a speed that was unthinkable a couple of years ago. But when it's time to ground those ideas, fine-tune the details, and get a specific visual finish, there are no shortcuts: in my case, I still need the traditional environment.
AI gets you started, but the true control over details remains in the traditional tools
In After Effects:
Timewarp filter: For the high-speed sequence of the 111 renders.
Circular mask and a wiggle expression: To simulate the organic movement of blowing a bubble.
Center blur: To blend the mouths that appeared too sharp in the raw AI output.
In Premiere:
Final cut, subtitles (always yellow on black for readability), and audio mixing with Essential Sound.
What I take away from this project:
That discarded material you beat yourself up over, thinking you've wasted your time, is sometimes the exact starting point for your next idea.
AI demands flexibility. You have to learn to rescue what’s useful from what it gives you, instead of forcing it to deliver exactly what you pictured in your head.
Traditional tools are still essential to provide that quality control in the fine details.
And now that you know the guts of the process, here is the final result. I hope you enjoy it.
Read Next
