From Idea to Ad: Rethinking Filmmaking Through AI
- 4 days ago
- 3 min read
Updated: 21 hours ago
Here’s my thought process behind making Just Another Day—an AI-generated ad that cost me less than RM100 to produce, including all the unused and failed shots along the way.
I made “Just Another Day” almost entirely inside a desktop window using Google Flow—no set, no crew, no traditional production pipeline. Just ideas, stitched together digitally.
And for the first time, I wasn’t thinking about what I couldn’t afford to shoot.
I’ve funded most of my short films myself, which means every idea had to pass a brutal filter: cost. Locations, scale, genre—they were all shaped by budget.
This changes that. Suddenly, the gap between imagination and execution feels… a lot smaller. One of the most important steps for me is creating a character technical sheet. It doesn’t need to be complex—in fact, I’ve found that simpler works better. As long as the core visual identity is clear, that’s enough.
This sheet acts as a reference point that I attach during generation, helping the AI maintain consistency across shots. Without it, characters can subtly drift—different facial features, proportions, or styling.
I actually learned this the hard way. I used the sheet during image creation in Nano Banana, but forgot to include it in some of the earlier video generations. That’s why you might notice slight inconsistencies in those shots. It’s a small detail, but it makes a big difference.
Another challenge I ran into was spatial control.
Directing mise-en-scène, character movement, even basic blocking—things that feel second nature on a physical set—became surprisingly difficult. Getting an actor to move from one point to another, deciding which hand they should use, or triggering a specific action isn’t as straightforward as it sounds.
What I realized is that the AI doesn’t really “understand” space the way a director does. There’s no true sense of spatial awareness, which makes precise framing and controlled movement quite limiting.
For now, the approach is less about directing every detail, and more about adapting—working with what the system can generate, rather than forcing it into something it can’t yet fully interpret.
Once I had a set of usable shots, I organized them into a “selects” collection ready for export into the editing stage, while the unusable takes went into a separate folder.
This step became more important than I expected. It gave me a clear overview of what I actually had—what shots were working, what was missing, and what needed to be regenerated.
More importantly, it helped manage AI credits. Unlike a traditional shoot where overshooting mainly costs time and storage, here it directly impacts cost.
In a way, it’s not that different from real-world production—you still need to know what you’re looking for. Otherwise, you risk either overshooting and wasting resources, or undershooting and not having enough to cut with.
For voice work, I usually rely on ElevenLabs to generate dialogue for my animated characters. But with Veo 3 now capable of generating dialogue as well, I started wondering—could I push it further?
Instead of just generating a voice, I tried prompting for a voice-over artist performing in a recording studio, as if I were directing an actual session.
The result was… unexpected. The system didn’t just produce audio—it created the presence of a “talent.”
The two voice-over performers at the end of the ad are entirely AI-generated. And seeing that play out still feels a bit surreal—like working with a voice actor who doesn’t physically exist, but somehow still shows up for the performance.
Comments