Do image-generating neural networks show brand preference? Should they?

Some discussion going on last week about the role of brands in text-to-image AI algorithms, specifically Dalle2 and MidJourney. And since no one asked me, here are my thoughts.

Right now, this discussion is completely academic. The people using these platforms are developers, artists, and enthusiasts. The commercial applications (rights management) of this imagery are just being hammered out, and the policies change week to week.

But let’s put all that aside and assume that 1) the parabolic pace of tech advancement continues and 2) in a couple of years, video game developers, content creators, VR and AR creators will be seamlessly using text-to-image imagery as a part of their workflow. Now the manner in which brands appear in outputs becomes very practical and relevant, as objects can appear in a generic form or as a brand. 

If a future VR user wants to materialize a Casino interior on Jupiter, what Casino brand is it? What brand of liquor is behind the bar? If a future game developer needs a 3D render of a fast food burger restaurant, is it going to be a McDonald’s or a Burger King?

As of now, it would be neither. After running tests for a few months, I can safely say that Dalle2, MidJourney, Disco Diffusion… currently none of them consistently produce branded imagery when a generic prompt is used. Fast food restaurant, Sneaker, Ketchup, Soda, etc… all generic outputs. Once in a while, you’ll get a hint of a swoosh on a sneaker, or a suggestion of a curvy glass soda bottle, but for the most part, the algorithms deliver generic items.

If you want a brand, you have to ask for it. Coke bottle, Heinz Ketchup, Nike Sneaker. Even the word brand… cola brand, ketchup brand, iconic sneaker brand etc. can work and produce branded imagery. 

So if a brand wants to appear as the default in AI generated imagery in the future, what are they to do?

Well, you have the AI’s input (what imagery the AI is trained on) and the output (what the AI produces in response to a prompt). Conceivably both of these could be influenced by brands. 

On the input side, a brand could in theory flood the internet with millions of images of their brand and hope that future models will train on their predominant imagery. 

Maybe not the best approach. 

Or, a brand could partner ($$$) with an AI platform (these people need to eat after all), hand over a library of brand imagery, and it would be folded into the model's training. 

Pardon the terrible oversimplification, but tweaking the inputs to the AI is, to me, analogous to using SEO to engineer how your brand appears organically in search. 

When asked during office hours if they would be open to a brand doing this with them, the enigmatic head of Mid Journey DavidH said (and I’m paraphrasing) “yeah sure I hadn’t really considered that but sure I suppose we’d do something like that. I gotta get outta here are we doing more questions?”

On the output side, I can definitely see the application of a more SEM approach, simply buying out keywords for image output, the branded output goes to the highest bidder. 

Maybe DavidH was being coy. Maybe MJ, Dalle, and Google’s upcoming Imagen, maybe they are all just building a platform for advertising revenue in the future. Certainly the current subscription and credit-based models are not real money makers and probably don’t cover much more than the GPU time needed to process the results. 

And of course, all of this depends on my above assumptions coming true. I think they will. What do you think?