From paying an image bank to our own workshop: how we created our visual resource generator with AI

In a burst of candor and generating significant media buzz, Anthropic CEO Dario Amodei warned us a few days ago of the impact artificial intelligence would have on entry-level office jobs. In an interview with Axios, the executive said that up to 50% of these types of positions could volatilize over the next five years.. The truth is that this is not an easy point of view to refute at this point in time, but in any case we cannot turn our backs on reality. We can, as individuals and as organizations, fight to adapt or entrench ourselves in a parallel reality. Best of luck to all, in any case.

What is explained in this article is not a leading use case or within the reach of few. It does have certain barriers to entry, but the truth is that they are not that high and it should be borne in mind that examples like this one will become increasingly common in all types of industries. Therefore, the article is destined not to age particularly well.

Do we still need image banks?

That question has been on my mind since September 2022. It was then when, after attending a conference, I was able to test some tools such as DALL-E or Midjourney. The truth is that it was the second one that I liked the most. It was and still is a terribly powerful generative artificial intelligence model that I have been lucky enough to work with quite a lot since then. Luckily it gave us the option to exit the Discord ecosystem ecosystem almost a year ago now.

Until not so long ago, every time we needed professional quality images, we had to either go through the cash register and buy a license or use more or less creative solutions. The playing field was a bit more complicated, but as a summary of the case we can leave it at that.

Graphic resource banks are practical, but they have several disadvantages. They are expensive for small and medium-sized companies. You always depend on other people’s styles and that is a limit to maintain a coherent visual identity. I would add that you also spend a lot of time browsing these banks and you don’t learn anything. It can be done better, really.

Why build a specialized GPT

Taking advantage of the environment of GPTs (Pre-trained Generative Transformers) environment offered by ChatGPT, we decided to develop internally at Smart Team a GPT whose main function is to stop being clients of any kind of graphic resource bank and whose secondary function is to allow the agency to be a seller on Adobe Stock (we would also be a seller on Shutterstock if this platform had understandable conditions regarding artificial intelligence).

The truth is that creating a GPT is not difficult if you like technology and have the desire to do it. The interface offered by Chat GPT is very intuitive and if you know what you want, you can get it in a fairly short period of time. I write down some parts of the process we follow:

1. Provide good instructions. Remember that you are talking to an LLM and that these instructions are generated in natural language. That means plain and simple that you must explain what you want the GPT to do and how you want it to do it. Therefore, it is worth making an effort to explain yourself clearly and precisely, but nothing more.
2. Define options to start conversations. In our case, the main function is to generate prompts for Midjourney from an image, but we have also incorporated editing and tagging functions as well as a test bench to monitor the evolution of GPT Chat itself as an image generator.
3. To provide the model with a knowledge base. We uploaded style guides and technical manuals from various databases, about 50 articles on the subject matter, our brandbook and all the recommendations published up to that time by Midjourney. Therefore, we provided the GPT with everything they needed to know about which images have sales potential, how to define the technical quality of the images and how the model we will use to generate them works.

The complete process for creating an AI imager

The truth is that, without forcing too much, in 60 minutes we can generate more than 60 good quality images and have them submitted to Adobe Stock for evaluation and approval if necessary. The next step would be to measure results for 4 hours in a row, but for the moment there is no time for the experiment. Maybe in August. Here is the process:

1. Design of a battery of prompts in our specialized GPT.
2. Generation, filtering, upscalling and downloading in Midjourney
3. Optimization of formats and measurements with an online online tool
4. Generation of headings and labels in our specialized GPT
5. Upload to Adobe Stock, apply titles, apply labels and submit for review

In short, if you dedicate an hour of your working day to create a gallery of graphic resources for your company’s website, you can contribute those same resources to a platform and hope that this “digital asset” ends up giving you some return. We can talk about how we think the Adobe Stock algorithm works another day, but the truth is that at the time of writing these lines we have managed to generate 4 downloads and we will have to fantasize about what would happen if we started to generate resource images compulsively and seriously. For the moment, we have a marginal yield but it is an asset not limited in time.

Results obtained

The impact has been positive and the process is here to stay and to be improved where appropriate. The main results are as follows:

1. We are no longer dependent on external banks. We create our own graphic resources and adapt them according to the needs of each project and each moment.
2. We have gone from buyers to suppliers. We have put images into circulation and have gone on to sell. We may try to scale this activity at some point and explore its potential, as well as know its limits.
3. We control quality and style. We do not depend on the creativity of third parties and have no reason to settle for something that does not suit us.
4. We create visual packages for our clients. We can and do deliver consistent image sets tailored to the specific need. For example, photo quality images for corporate websites

You don’t have to be a big technology company to create your own tools with artificial intelligence that have a direct impact on your business. I insist that there are still barriers to entry, but carrying out a process like the one described here is not that far from what most companies can do if they set their minds to it. At the time of writing these lines, I believe that any communication agency with a certain turnover should consider the possibility of no longer depending on graphic resource banks.