5 Insights From My CollXbs Experiment // BrXnd Dispatch vol. 019
A call for cool AI work + A few (hopefully) interesting thoughts from watching thousands of AI-generated collabs come by my screen.
You’re getting this email as a subscriber to the BrXnd Dispatch, a (roughly) bi-weekly email at the intersection of brands and AI. On May 16th, we’re holding the BrXnd Marketing X AI Conference. Tickets have gone much faster than expected, and the waitlist is now open. Add your name to the waitlist to be informed when more tickets are available. With that said, sponsor tickets are still available, so if you’d like to sponsor the event, please be in touch.
I was traveling last week, so I missed getting one of these published. That means I’ve got a few housekeeping items to share:
During breaks at the event in May, I would like to show off some cool examples of creative AI work from agencies, tech companies, and individuals. My thought is that I’ll make a big loop of all the videos and stick a QR code in the bottom corner that will have more info on the creator. If you’re interested in having your work included in that, I’ve set up a submission form. You can either include a video link or just upload it directly.
I’m looking for someone interested in sponsoring after-conference drinks. If that’s something you’d be up for, let me know.
Lots of info about speakers and sessions are up on the BrXnd Marketing X AI Conference site.
Alright, onto the content. Last week I was in California giving a presentation that builds on some of the things I wrote in my post about building intuition for AI through tinkering. Part of that talk is pulling out some of the more interesting insights from watching thousands of brand collabs flow through my CollXbs tool. If you aren’t familiar with CollXbs, the idea is simple: you put in two brands and choose an item, and out the other end pops an AI-generated brand collab. While many excellent examples of AI-generated collabs are floating around, what makes this different is a) anyone can do it without signing on to Discord or anything else, and b) it’s a reasonably unique pipeline that generates the image. Here’s what it looks like:
I feed the two brands and item to a model I fine-tuned on hundreds of actual collab announcements. This gives me pretty realistic-sounding marketing text to go with the AI-generated collab.
I fine-tuned another model from that marketing copy that extracts an aesthetic description of the item.
Next, I take that aesthetic description and transform it into an image prompt. This utilizes another fine-tuned model.
Finally, I feed that prompt to Dall-E. I wish I could use Midjourney, but there’s no API available yet. Certain items and brands get a bit of pre-processing. For instance, sneaker prompts get appended with “8k. ultra realistic. commercial photoshoot.”
From there, everything processes, the collab is posted, and an email is sent out. It takes a few minutes mostly because I built it a while ago in a slightly sub-optimal way, but the waiting is kinda fun, I think.
And now, after seeing thousands of AI-generated collabs, here are five insights I pulled out for the presentation that might be of interest.
First and most importantly: good brands make better images. This is really what sent me fully down the rabbit hole of marketing and AI. While that statement seems obvious at first, if you think a bit harder, it’s not obvious at all that a machine learning model would come to the same conclusion as we do about what makes a good brand. And yet, consistently, the best brands kick out the best images. So why would that be? My best guess is that the things that make brands great—practices like consistency and distinctiveness—are particularly well-suited to the ways these models pick up on patterns during training. So in a way, it’s a perfect match: brand-building is about creating patterns that lodge themselves in people’s brains, and machine learning is about picking up on patterns that are visible and invisible to humans.
Nearly every time I talk to people about how interesting it is that good brands make better images, they respond that it probably has to do with scale. That’s not unreasonable: the best brands are almost definitely better represented in the corpus that is the internet than their smaller, less established counterparts. But what I’ve found in the outputs is that it’s not unusual for strong small brands to also shine in their imagery, leading me to believe it’s not just scale that the model is learning. Brands like Aimé Leon Dore and Ghostly are nowhere near the size of Nike or Coca-Cola, but they still come out looking great. Conversely, large weaker brands don’t come out nearly as good as their stronger counterparts. (I’m sparing those names for now, but if you doubt me, click around the brand list.
A huge thank you to the many supportive sponsors who are making the event and my work possible. If you would like to sponsor, please be in touch.
A huge thank you to our sponsors at The Brandtech Group, LinkedIn, Redscout, EZNewswire, Nova, Horizon Big, Otherward, Persistent Productions, McKinney, and The Imaginarium. If you’re interested in sponsoring, please be in touch.
One of my favorite findings—and something I’m working on building on more—is the idea that this method of generating collab images is actually a fascinating way to measure the relative strength of brands. The example this is most apparent for is Hermès. If you try to generate a scarf collab between Hermès and another brand, Hermès almost always completely overpowers. If you look at the scarf on the left, meant to be a collab between the FT and Hermès, the other brand is almost wholly lost. Click around the Hermès page on CollXbs, and you’ll see much of the same. Unless the brand being collaborated with has an aesthetic that matches the power of Hermès. The example on the right is The Simpsons, which you wouldn’t typically think of as a brand, but clearly is. Another favorite of mine is this Grateful Dead x Hermès scarf:
I really love that scarf.
Another fun one is that while we don’t typically think of art direction as part of a brand’s identity, it absolutely is, and for brands with a strong art direction aesthetic, like Supreme, you can see that the model has picked it up.
Finally, and maybe my favorite, is how sometimes, when you ask for something, like a sneaker, it puts a swoosh on it regardless of the brand you chose. The example on the right was a request for a Dunkin’ x Saucony sneaker, which, despite the brands, comes out branded as a Nike. At first, this seems like an error, and I guess you could say that’s what it is, but I’d argue it’s a good indication of just how well these models understand the world of brands. For most people, sneakers = Nike. I don’t think that’s a particularly controversial take. They’re by far the largest brand in the category, and, maybe more importantly, everyone has been biting their style for years.
The challenge with having a brand leader who is also a style setter means that others often end up effectively creating marketing for them. One of my favorite stories from my early advertising career was working with a consumer electronics company desperate to appear cool. They spent a lot of money making a new TV commercial featuring snowboarders and other “edgy” stuff. When the commercial went out for testing, the consumers loved it, and when the final question of unaided awareness came back, they confidently said it came from Sony, which was most definitely not the brand that had produced the ad. In the early 2000s, Sony dominated the market and had some of the best ads in the world. Trying to make something that played in their aesthetic wheelhouse only strengthened their position. Millions of dollars were spent to create an advertisement for their biggest competitor.
I think that’s what is happening with Nike. If you think about how the diffusion process works—noise is reduced until the image roughly matches the prompt—it makes sense that the output often ends up being a Nike when you ask for a sneaker: that’s probably what a human would think as well.
That’s it for this week. Please don’t forget to submit a video of AI creative work if you’re interested.
See you in a few weeks!
PS - If you haven’t joined the BrXnd.ai Discord yet and want to talk with other marketers about AI, come on over.