A friend sent me an IBM article about LoRA and told me I was "a beast" at it. It made me laugh, because my first thought was "What even is that? I don't even know what it is, a beast... I can't call myself a beast at that"... But I read the whole thing and, to my surprise, I got it faster than I expected. And I think that's because it's something I'm living right now with a campaign.
So I'm going to explain it the way I understand it, from design, without so much math. Maybe you're a beast at it too and you don't know it yet.
The big AI models, the ones that write and the ones that make images, get trained once on a huge amount of information. They come out knowing a little about everything, but about your stuff, they really know nothing: not your brand, not your character, not how you work.
Before, to teach a model something like that, you had to retrain the whole thing, and that's expensive, slow, and pretty much out of reach. To give you an idea, GPT-3 has 175 billion parameters, which are like the internal settings the model moves around when it learns. "Retraining" it means moving all of those settings... ALL of them.
LoRA (Low-Rank Adaptation) does something different. It leaves the original model alone, untouched, and puts a layer on top of it, and that layer is the only thing that gets trained. It learns only what the model is missing to do what you need. The article mentions that with GPT-3, what had to be trained went down about 10,000 times, and the memory it needed dropped to a third.
The "low-rank" part is the technical bit, and in a few words it means that adjustment fits in very little information, because these models have a lot to spare. If you're curious about the matrices, the IBM article explains it.
When I read it, the first thing I thought was: "Wait, this is a brand book."
When an agency hires a good designer, they don't send them back to school. That designer already knows composition, type, color... What you give them is the brand guidelines and a few reference pieces, and with that they adapt what they already know to that brand. The designer is the same, the only new thing is the "layer" you put on top.
Something else I liked: since the base model doesn't change, you can have several layers and switch them depending on the job, one for each client. Just like the freelancer who has a folder for every client with their guidelines, their fonts and their colors, and opens whichever one fits whoever just emailed. (Yep, that's me.)
These past few weeks I've been working on a campaign for Aerosol with AI-made images. It's all still in drafts, nothing has been published yet... And that's when I realized that what I've been doing for a while now looks a lot like what LoRA is trying to solve.
This didn't start now. I've spent four years experimenting with AI, feeding and refining the agents. And a little over a year ago I ran a workshop with the Aerosol team to define the brand: the strategy, how it talks, the art direction, the visual language, the logo style, the palette... I've been giving them all of that too. That's why what comes out today already starts from our brand. There are still changes and rounds, but on top of something that's been built over time. If you ask a generic model for "a person holding a phone", you get a stock photo that could belong to anyone. And some details it just makes up.
What worked for me was building a full process with tools, from start to finish. The photo gets generated with the phone screen in green, the real app screenshot goes on top in perspective, the person gets cut out so the type can sit behind them, and the composition gets put together with our font and our system. The tools I've been setting up do all of that. I don't cut out or assemble anything by hand, what I do is direct: I look at the result, ask for this or that to change, and keep going until the piece looks the way I want it.
Now, to be honest, we haven't trained a LoRA yet. The tools we use work with the documents and reference images we give them, and so far that's been enough. But we already have the foundation: a clear brand, the material, and a process that's ready for that step. My plan is to keep learning and, in the near future, start experimenting with training one, for example so a character comes out exactly the same across fifty pieces without me having to ask for changes piece by piece. When I do, I'll tell you about it.
LoRA has a cost, and the article mentions it: when you shrink things that much, something gets lost, and you have to test how much. From the design side, I'd add another one: a LoRA learns from what you give it. If you train it on mediocre material, it'll give you back mediocre material, faster and in bulk. And if you give it a well-thought-out system, that's what it'll multiply.
So the hard part is the same as always: having a clear brand. Knowing what it is, what it isn't, and what it doesn't need. That little layer you put on top of the model only works if someone knows what goes inside it.
I know there's a lot of criticism of AI, and of those of us who use it. I get part of it, because generic content exists, and you can really tell when there's no brand behind the model. But, for better or worse, that's where the world is going, and the professional who stays behind... That's exactly where they stay: behind.
In my "company", I'm the only person. And using AI this way has saved me a ton of time, sped up my processes and given me support I wouldn't have on my own. For me it's been a blessing, truly. And the more I learn to use it, the better what I can offer my clients. It keeps me up to date, prepared, and I feel like it puts me ten steps ahead.
And not just in design. It even helped me start understanding taxes and doing my own books, with AI as my assistant. Along the way I realized the person I was paying to do it wasn't doing it the best way. Now I can do it myself, with less room for error and one less monthly expense.
And one more thing that almost never comes up in that conversation. There are teams that can't pay for a big production, with a photographer, a colorist, a studio, talent and everything that comes with it. For a lot of them, the real choice isn't between that production and AI, it's between AI and not doing the campaign at all. That's where people like me come in, people who know how to handle these tools, to help them reach their goals with what they have.
And what I see, at least for now, is that AI needs exactly what we do so it doesn't come out generic. That's why, for me, this brings more curiosity than fear.
A beast? Not yet. But I already understand where the layer goes and what goes inside it. The rest is practice.