Hip Hip! Hooray! OK. You get the idea...

Here at DoltHub, you may have noticed we've recently been changing our marketing copy. We realized AI agents need branches in order to work safely and effectively. Dolt is the only SQL database with branches. Thus, Dolt is the database for agents. Previously, our marketing focused on what Dolt is: Git for data. Now, we're excited to have a use case that really exhibits what Dolt is for: agents. We think agentic workflows are going to be huge and Dolt can be the backbone for every one of them that's not code.

Creating good marketing copy is hard, specifically the imagery associated with the copy. The good news is that AI image generation is getting really good. It's not good enough to make production images but it is good enough to help with the creative process. I use it to send more fleshed out ideas to our designers. This article walks through two examples.

AI Image Generation Alternatives

All the major generative AI websites now support image generation natively. I've used it on ChatGPT, Grok, and Google Gemini. My favorite for speed and quality is Google Imagen that comes as part of Gemini. I'm not sure if is free or paid because we use Google Workspace for our email and productivity software. It is free as part of that package.

When image generation first came out, you basically had to one shot the image. The tool was incapable of making subtle edits to a generated image. The image generator was not very good at visually matching the styles of input images either so the images generated were of a random style or worse, "AI-style".

Those problem have been fixed. Now you can iterate on an image to get it closer to what you are looking for. The image generator can closely match image styles and even copy individual objects in source images. These changes allow you to get far closer to an image your proud of.

As you can see, image generation and editing is still not perfect. Some of those robots still are only heads and there are the same number of them.

Having a tool that can generate images based on concepts allows you to quickly test your marketing ideas. Do you like the way the image conceived from your concept looks? Yes. Let's get a designer involved. No. Back to the drawing board. You can turn around concepts in minutes.

Prototype Only

Image generation is still not good enough to make production quality images. So, my process is to use it for getting tangible feedback on my marketing ideas. Think of an image idea. Create a detailed prompt explaining what I want. Have the AI generate the concept. Iterate. Once I have an AI generated image that I think will work, I send the prompt and image to our human designers. The designers usually take a few days to produce a production quality image from the concept.

Let's look at a couple examples of this process in action. One of the images is in production and the other should be going on the DoltHub homepage this week in a new section.

The Database for Agents

We added a new featured image to the homepage that matches our "The Database for Agents" tagline a couple months ago. We love the image. It is a perfect example of how AI image generation can enhance the creative process. I wanted a tree sprouting out of a Dolt database cylinder like a potted plant. I wanted robots to represent agents working on the branches.

This was the prompt:

A branch like tree sprouting from a Dolt database cylinder with robot icons as the leaves.
<Attached Git for Data homepage image> 

This was the best result after a few tries a few months ago.

You can see it's ok but it's not matching styles are objects as well as the newer models. I sent this over to our designers and they came back with this beautiful image.

I like how the human designers stylized the tree into more of a commit graph like structure. I also like how the agents are sitting and standing on the branches as if they are working on them. This human touch definitely made the image what it is.

Would we have achieved the result with only the prompt above to the human designers? I'm not sure but I'm very happy with the result we achieved using a combination of AI and human designers.

Agents Need Tests

We're adding new "Agents Need Tests" section to the homepage. The image is ready and we followed a similar process. I had an idea to have a side-by-side medical scan inspired image of a good and bad database.

This was the prompt:

I would like an image in this style. I want two dolt database cylinders like the one at the bottom of this image. The one on the left has a crack in it and it is being scanned by a medical looking testing machine. The output of the machine says "compromised" and "tests failed". A red eyed agent stands on top of the cylinder looking menacing. The one on the right is also being checked by a similar looking medical test device but it looks perfect like the one in the attached image. A white agent agent stands proudly on top of that database. The output of the machine is a green "all clear" or "tests passed".
<Attached agents on branches image above>

This was the best result after a few tries.

A little busy and a bit out of style but it captured the essence of what I was going for. I sent this over to our designers and they came back with this image after a couple of iterations.

I like how the human designers removed a lot of unnecessary clutter from the AI concept and made the image styles match exactly. Again, I'm really satisfied with the result and I think AI image generation helped the creative process.

Conclusion

AI image generation is good now. You can start integrating it into your creative process. We are using it to help bridge the gap from idea to final production marketing images. Do you have other AI tools you think we should try? Stop by our Discord and please let us know. I'll write about them if they are good.