AI Projects

This one's an experiment I keep coming back to: Stable Diffusion running in real time, so I can art direct a product scene the way you'd direct a set. Type kitchen, we're in a kitchen. Office. Garage. Park. Green headphones, red headphones. The scene follows the direction as fast as I can give it.

Same rule as everything else here: it's all about minimizing hallucination. The product rides on a locked 3D model, and presets keep it accurate while everything around it stays flexible. A work in progress, but you can see where this goes: live art direction for prototyping, moodboards in minutes instead of days, and testing a scene before you ever book a shoot.

This one fills the gaps around a shoot, it doesn't replace one. You need an angle that never got framed, or the colorway that didn't exist at the last shoot. Most AI tools take one photo and ask the model to imagine the rest, and on a $300 product where the logo and texture matter, it guesses wrong. So I lock the product first: reference photos, or an exact render from a 3D model. AI only builds the world around it.

This isn't one photo in, a thousand ads out, spray and pray. And no AI people, ever. That's a design decision, not a limitation. Products and lifestyle scenes only.

First drafts from AI are generic. Everyone who's used it knows this. The value is the system around it. I pull what competitors are actually running from the Ad Library, sorted by how long each ad has been live, because nobody keeps paying for a loser. I scrape their landing pages so I can see the whole funnel message, and I check Reddit for the pain points nobody in the category is answering. That becomes a written test plan I sign off on. I keep calling the clusters themes, but really they're formulas that are already working out there.

Then copy runs through eight scoring agents until everything clears 90. Three to six rounds, usually. That loop is what filters out the slop. These need to be polished and on brand, and I want to lean into the data of what's actually working, because I don't want to waste the company's money. Beautiful ads don't necessarily convert. The test is: is it pretty, and did it work?

This one is for brands that are nervous about AI imagery, and honestly, I get it. The pixels here are ray-traced from a 3D model, not generated. No hallucinated labels, nothing drifts. AI just runs the production line: scene setup, batching, QA.

If you have 50 SKUs and need the front, side and back of each one in identical light, this turns a quarter's project into an overnight render job on cloud GPUs. When a new product launches, you add one line to the job spec and run it again. And 3D has gotten so good I'd pick it over photography for a lot of e-com even if AI didn't exist.

This agent sits between the media buyer and the creative team, which is a seat that usually doesn't exist. I've spent most of my career being the needle between performance and brand, and this runs that seat weekly. It reads the meeting transcripts, it reads the ad account, and on Monday morning it tells us what to build, what to kill, and what's due.

The part I like most is the memory. Send it an ad concept over Telegram and it answers with history: this theme ran last holiday at the deeper discount and converted, the shallow one didn't. I think creative teams need to be attached to the data. Not drowned in it. Attached to it.

I built the first version of this at Momentous and it's still the one I use most. A shoot happens for one hero video, but the transcript has lines in it a copywriter couldn't write, said by a real person with a face and a name. I don't like reading transcripts, so I made the system do it. Some of the best hooks I've ever run came out of a transcript nobody had read twice.

And it's an archive. Come Black Friday, you can ask if any ambassador ever talked about a product, and the answer comes back with timestamps, from a shoot everyone forgot about. One rule: never put words in anyone's mouth. Anything published under a speaker's name is their verbatim words, and they sign off first.