Make ad variants with AI: create an image first, then turn it into video

… I'm working with creating ads for some of my clients. And we like do 99% of it with AI by now. So it's always our go to that we want this to work with AI, like some formats is probably wouldn't make a difference, like from a cost perspective and execution, like you could, like if you have like creators you're working with, you can just request content off of them. It's going to be incredibly cheap if you work on an ongoing basis. I think

like using AI, hoping it will be cheaper, that's not where we're at yet. Like if you kind of have to do a few generations, it always like racks up a bill as well. Someone still needs to do it, learn the tools potentially. I wouldn't do it from a cost savings perspective, but to me, if we can do it in AI, that means variables become very controllable. It's not me writing briefs, trying to figure out how a creator acts the way I want them to, but we can learn how to do it with AI and then kind of really invest into making sure we control these variables and can

test them. Right now we're in this kind of middle ground where both ways work but I would say right now AI is a very plausible investment for me because in half a year it's gonna be cheap, it's gonna be like much more testable, you have everything encoded in like a machine and then you, I mean everyone is dreaming of this being a closed loop of course like when your AI tools are creating these ads and you also connect your meta ad account …

… like certain training exercises or so, then it might be another model. So you have to really play around based on the content that you're trying to create. And then different models will have different training data and some stuff is just going to be hard to create. On the image side, it's mainly GPT Image 2, which recently came out that has been pretty, pretty strong, especially when it comes to like keeping text consistent and well working

while Google's nano banana model has been especially fine print stuff. It's not so great, but both models generate very, very strong imagery. And usually the process is you generate images first to make sure that you've created a scene that you feel like is worthy of my ad. It looks authentic. It's not like AI plastic look. You kind of came up with something that you want to use in the style of ad you're doing. And then you use that as the reference image within the video model

because it's going to be much easier to get it right with an image model and it's much cheaper. So you can, for like a few credits, create tons of images until you have the perfect one and then only make it into video from there, meaning your C-Dance or your VO is going to use that exact scene you created and then just add on to it. While if you prompt a video model directly, then it will take a lot of time until you feel like, okay, this is there and it will spend a lot of credits. So rather do it that way. The other thing interesting is there are these kind of node-based tools like Weevy and Flora.

Weevy was recently acquired by Figma, which allow you to then basically build workflows. So as soon as you have like a style where it's like, okay, we found that text walls, for example, work. And we always want to create text walls that consist of two scenes, an app demo in the middle, and it's going to be text block on top. You can pretty much automate that in something like Weevy. It plugs in with all the models as well. you define how you kind of what's the prompt structure for each of the scenes and can think about then feeding variables for example so that you can kind of then create that ad format based

of kind of a node-based workflow and then iterate quickly because as soon as i said before like we had ads that we made 400 500 iterations on as soon as the format is established you're looking to keep it alive and you're looking to improve it you want it to like scale into different placements and really make sure that you can spend on it for a long, long time. And hence, you're going to be making a lot of it. And that's where too like this can …