A Guide to Creating Ad Visuals with AI

Let's put the prompts and the workflow to the test right away by creating a banner for a social campaign for a hotel in Venice.

The brief

To recap: I need to decide what I want to get. Which elements I want in it, what the information hierarchy should be (meaning what I want the user to see first), what the copy should say and how much space it should take up. What action the user should take, the colors, the contrast, which kind of font to use.

Let's do it by building an ad campaign for a hypothetical boutique hotel in Venice that wants to sell romantic packages for couples.

Creative Brief: Advertising Images for a Boutique Hotel in Venice (Gemini 2.5 Flash)

We have a brief, we've got a great guide, and now it's time to find some ideas for our campaign

Phase 2 - Research

Before you start firing off random prompts, it's essential to spend time on visual and conceptual research. This phase helps you pin down the art direction and enrich the brief with concrete ideas. The main tools are the classics of the creative world: reference imagery and mood boards.

The mood board. Real images and databases of images generated by others

Before you start firing off random prompts, it's essential to spend time on visual and conceptual research. This phase helps you pin down the art direction and enrich the brief with concrete ideas. The main tools are the classics of the creative world: reference imagery and mood boards.

Mood board: a collection of visual elements – images, sketches, colors, maybe keywords – put together to evoke the atmosphere or style you're after. In practice, a collage that shows the look of a project at a glance​.

In our case, building a mood board at the start helps you stay consistent: you can include images you found online or in the company portfolio that capture the right tone for the AI images you want. For a winter campaign, for example, you might collect photos of snowy landscapes, vintage Christmas graphics, illustrations in cool colors and so on. The mood board will be your style North Star every time you judge the generated results.

Reference images: beyond the overall mood board, it's worth picking specific images as references for particular details. Say you want to generate a render of a new product: you could take the lighting from an existing professional photo as a reference, or a model's pose in a certain shot, or the line style of a famous illustration.

You can use these references by eye (that is, looking at them and then describing them in the prompt: "dramatic lighting like in Blade Runner") or directly through AI. Many generators let you feed in an image along with the text (the image-to-image technique): Midjourney, for example, lets you put an image URL in front of your text prompt, shaping the composition around it; other tools like Stable Diffusion offer advanced controls (ControlNet, etc.) to make the output resemble a given image.

You can also use dedicated tools: for example, the CLIP Interrogator on HuggingFace is a model that takes an image you give it and generates a possible descriptive prompt​, which helps you figure out which keywords to use to get something similar from a text model.

AI communities and databases: online AI art communities are a gold mine of inspiration. Join forums and Discord servers full of creators and you can see what others are making, pick up tricks and get ideas. The official Midjourney Discord, for example, is a great place to browse thousands of images generated every day and read the prompts behind them (unless users keep them private).

Beyond interactive communities, there are ready-made prompt galleries and AI image banks. These are extremely useful: by studying other people's prompts you learn new keywords and structures. You might discover, for instance, that images you like use terms like "cinematic lighting" or "ultra-detailed, 8k". Write them down! A good habit is to build a small personal database of interesting prompts you find online: save them in a document, maybe sorted by category (e.g. portrait prompts, fantasy landscape prompts, etc.), so you can adapt them to your needs. Reverse-engineering other users' prompts is great practice: take a complex prompt and break it down, and try to understand what each fragment does (there's the subject part, then maybe an art style, then technical parameters). With practice, you'll start to recognize common patterns that work.

Vision Models

To read your mood board better and find the keywords that pin down the style, you can use vision models. Put simply, AIs that "see."

Vision models are AIs that specialize in analyzing and interpreting visual information, which includes, among other things, classifying images and generating text descriptions.

To pull the "aesthetic reference" out of an image, you use specific models or tools (often based on CLIP, like the "CLIP Interrogators," or models like Llava/BLIP that can write detailed descriptions). They analyze the image and produce a text description that includes style elements: type of art (e.g. "watercolor," "photorealism"), an artist's style (e.g. "Van Gogh style"), lighting (e.g. "cinematic lighting"), mood, technique, etc.

Then, as I said, you can take that text description (keywords, phrases) and drop it into an image model's prompt, adding the descriptive terms to the subject you want (e.g. "a sleeping cat, pastel watercolor style, soft lighting"). That guides the generator to create a new image that captures the look of the original reference.

We'll look at this in detail in the next phase. In the meantime, here's a solid prompt for analyzing images for our purposes:

Visual Brief

And here's the result of our research, which we'll use in the production phase:

Phase 3 - Production

Now we get to the heart of the guide: producing the images.
For this test I picked the state of the art as of January 2026. To see which models the pros like best, check LMArena https://lmarena.ai/leaderboard

The stars here are: seedream v4.5, flux-2-pro, nano-banana-pro and Reve. An honorable mention goes to Midjourney: still relevant but on the way down, which is why I wanted to give this guide a fresh coat of paint. There's a quality new entry, OpenAI's GPT Image 1.5, but I'm not convinced by how it handles advertising images, so it isn't very relevant to this guide.

Now we put the plan into practice. The process breaks down into a few key sub-steps: choosing the right tool and settings, writing the prompts (starting simple and then refining), possibly using language models to improve them, and finally generating and iterating until you get a result you're happy with.

We have plenty of options, as you saw earlier. I chose these two tools because they're easy to use, the results are good, and the differences between them show nicely what it means to treat the tool as a creative choice.

Keep in mind that every model reacts to prompts differently, so each one needs its own adjustments. That detail adds another layer of complexity, and of fascination, to our trip through AI image generation.

As I said at the start, you don't create images, you choose them. Or rather, you set a direction and then pick the one that fits best. How do you choose? You need to know the basics of advertising and visuals. Through the prompts, I'll try to bring in all the theory you need to choose well

Basic Prompts

Let's start with basic prompts, which give the model maximum freedom to generate the image probabilistically, and work our way up to more granular control

Basic Prompts (also called ZeroShot) = Leave it to chance
Basic Prompts are simple and direct and give the AI a general direction on what you want to see.

For this guide I want to do an exercise and create a campaign for a hotel in Venice. I need an image:

Seedream v4.5

Flux 2 Pro

Nano Banana Pro

Reve

They're pretty, but there's nothing in them that helps me win the war for attention: I need to find a way to connect with the needs of someone who, right then, was thinking about Venice, a vacation, or something that might make their ears perk up.

Do you know how Berlusconi got his start as a businessman (at least the version he told us)? When he was trying to sell apartments, he furnished them, so visitors could immediately put themselves in the picture and see themselves living there.

That's because visual information is processed almost instantly, and in parallel, by our visual cortex, a skill honed by evolution. Reading text, by contrast, requires sequential decoding (letters, words, meaning), which is inherently slower.

As a result, the message of an image lands far more immediately.

Studies like Simon Thorpe's (for example, Thorpe, Fize & Marlot, 1996 in Nature) showed that the brain can process and categorize complex visual scenes in a very short time (on the order of 100-150 milliseconds).

National Science Review (2024) - Diffusion Models and Stochastic Image Generation; Thorpe, Fize & Marlot (1996) 'Speed of processing in the human visual system', Nature.

So how do we find the right image? We have to design it, drawing the expertise out of language models. Every serious project starts with a brief; we already saw in phase 1 how to create one, and I followed exactly those instructions to build my campaign for the Venice hotel.

Structured Prompts

Structured Prompts are more complex and can include extra details or specific parameters. They're useful when you want more specific output. An example of an Advanced Prompt could be:

Seedream v4.5

Flux 2 Pro

Nano Banana Pro

Reve

While Basic Prompts give the AI some room to interpret, Structured Prompts tend to narrow that room to produce results closer to the vision you have in mind.

Advanced Prompt

Image prompts: more skill than art or science

Prompts, as you know, are the main way we talk to AI. You give it instructions to process and you get back what you asked for, in this case an image. But owning a star chef's pans doesn't make you a star chef. A well-written prompt can make the difference between a pretty image and an image that makes sense. Where's the difference? In the expertise.

Later in the lesson you'll see exactly what I mean.

Let's start with a few ground rules.

Be specific: The more details you give, the more guidance the AI has on what to generate. But stay concise and try to avoid filler words.

Let's look at a practical example.

Midjourney has a tool that shows you which tokens it considers and how much weight it gives each one when it creates images.

For example, let's put a bunch of words into the prompt we got from the vision model in the research phase:

We get

Seedream v4.5

Flux 2 pro

Nano Banana Pro

Reve

If we ask the AI, using Midjourney's tool, how it reads the prompt (the /shorter command), we find that most of the words (tokens) have no effect at all on what the tool produces

Cutting the "useless" words, we can write:

and we get a similar image with a similar mood from far fewer words. Of course, zero control, zero charm. Let's keep going.

Along the way, my experiments turned up a few good practices:

• Experiment with synonyms: sometimes small changes in wording can produce very different results.

• Think about word order: The order of the words in your prompt can change the output.

• Feedback loop: after generating an image, use what you learned to refine and improve your next prompts.

• Experiment and Observe: one of the best ways to find the right prompt is to experiment. Generate several images from variations of a prompt and watch how small changes affect the output.

Prompt Hacking

Trash in, trash out.

As I mentioned, to get images that make sense, you need to know your subject.

It's not like starting from a blank canvas and painting what you want, or walking around with a camera and taking random shots. Whatever prompt you give the machine, you'll get an image; but to get the image you want, one that fits inside a campaign, you need to know quite a few things.

If you want to create photographs, you need to know about lighting, photographers, angles, cameras, lenses and so on. If it's a painting or an illustration, you need to know painting techniques, art movements, artists, and so on. Now let's see what all that looks like in practice.

Creating photorealistic images

One of the most fascinating uses of AI image tools is their ability to produce images that look incredibly real. That ability has huge potential in marketing, where authenticity and detail can make a strong impression on an audience.

Here is the prompt, piece by piece:

Subject: a couple in their 30s
Composition: in a gondola
Aesthetic reference: spring 1980
Mood: happiness
Light: direct Sunlight
Shot: Wide Shot
Gear: Minolta X-570, Kodak Professional Portra 400 Film, Kiron 75-150mm f/3.8 lens

So the template will be:

Let's change just a few elements and see what happens. In this exercise we're going for a more elegant look by changing the "mood."

The prompt could change like this:

By changing only the mood [charmed] and the light [moody light], we got a completely different atmosphere.

Bringing artists into your prompts
One shortcut to a particular effect can be to draw on specific photographers. Whether that's stealing or inspiration, I'll let you decide after you've seen this experiment.

To get closer to the final effect we want, we'll pick the artist suggested in the brief: Annie Leibovitz.

Annie Leibovitz is an American portrait photographer best known for her portraits, especially of celebrities, often showing her subjects in intimate settings and poses (wikipedia).

Here's one of her fashion shots.

Credit: https://www.manamediagroup.com/nike-annie-leibovitz

What we have seen so far is that the more we narrow the field, the more control we have. Now we have two options: copy or draw inspiration.

let's add a piece to our prompt template

If we want to copy her creativity and let the machine do the work on its own, we go back to the basic ZeroShot prompt, like this:

Welcome to the world of fashion. Now let's try drawing inspiration instead and setting the mood ourselves through the lenses and the aesthetic references.

let's add a piece to our prompt template [artist or art style]

Right away we're dropped into a fashion look thanks to the artist we named, but with the mood we set at the start, driven by the aesthetic reference (Spring 1980), the lighting and the gear, which makes it feel much less like a "studio" image.

My personal takeaway from this experiment is that you can use artists, not so much to "steal" their style, but to set the scene better and better capture customs, habits and body language.

With a few changes that come from specific expertise, we went from a flat "family photo" to a far more emotionally intense shot.

Is it enough to win the war for attention? I don't think so.

Let's change the subject to match the brief:

Visual Idea: evocative images that suggest intimacy and secrecy. Key example: the hint of a gondola gliding silently past, seen from a half-open hotel window. Focus on atmosphere, details, the play of light and shadow, suggesting more than openly showing.

We've got our composition, and it all starts to make sense: there's suggestion, there's body language.

Let's go further and add all the references we found in phase 1, the kickoff, and in phase 2, research, while keeping our prompt structure:

Drumroll, please...

I'd say we've reached a more than respectable result. Of course it isn't 100% perfect: it isn't Venice, the hotel shown doesn't exist, and so a real professional shoot would be better. But that isn't always possible.

Before

After

You'll also notice plenty of mistakes in the generations. That's because I didn't want to cherry-pick or iterate: I wanted to show the current limits and potential of the models in 2026. The workflow can stay the same for the next 5 years or more, while the models will keep changing.

With this workflow I wanted to show the pieces you need to reach a result that's "under control" and well curated.

Art styles

Just as with photos, the models can produce images in specific art styles. That range lets you explore and recreate traditional styles, and also mix them or create completely new blends.

Prompt examples for art styles

Let's keep our starting example, but this time give it a painterly style like Impressionism, swapping the artist and removing everything related to photography:

Cubism

Aesthetic References

In advertising, aesthetic references are defined sets of visual traits – like colors, lighting, composition, texture – that evoke a specific style, mood or atmosphere. They work like a coded visual language, which is especially useful now with Artificial Intelligence: the terms that describe these aesthetics go straight into your prompts to guide the AI toward images with that exact look and feel.

Their power lies in their ability to instantly call up a historical period, suggest a type of product or service (as in the "Abstract Tech" example, where keywords like "glowing lines" and "futuristic designs" in the prompt produce images suited to tech products or sci-fi settings), or create an image that resonates with a specific audience (for example, a prompt with "minimalist" references for an audience that appreciates understated elegance). In short, using the right aesthetic references in your prompt lets you create images that clearly communicate the context you want.

Below are a few examples to experiment with.

Abstract Tech

Acid Pixie

Acid Wave

Acubi

Adventurecore

Afrofuturism

Afropunk

Ancient Egypt

Androgynous

Phase 4 - Distribution

Outpainting

Outpainting is a generative AI technique for extending an image beyond its original edges, adding content that fits the existing scene. In practice, the AI analyzes the pixels, style and composition near the edges of the image you provide and "imagines" how to continue the scene into the empty space around it, which you define. It works by giving the AI the starting image and marking the outside area to "fill." This is especially useful for creating different versions or formats of the same image: you can start from a smaller, detailed format (for example, a square 1:1) and use outpainting to expand it step by step (for example to 16:9 for a banner or 9:16 for a story), adding background or context smoothly without generating a brand-new image for each format.

Upscale

Upscaling is the process of increasing an image's resolution (meaning the number of pixels that make it up) using AI algorithms. Traditional methods simply enlarge the existing pixels, which often leaves images blurry or pixelated. AI upscaling tools, by contrast, are trained on millions of images to learn to recognize patterns, shapes and textures. When they enlarge a low-resolution image, they don't just duplicate pixels: they add plausible details and rebuild the missing information, trying to "guess" what the image would look like if it had been created at high resolution in the first place. The result is a bigger, sharper, more detailed image, ready for print or high-definition screens.

And that wraps up our workflow. More guides will follow to help you "hack" your prompts even further.

Stay tuned to Marketing Hackers!

If you found this useful and you like the project, help me grow it. Share it with your colleagues and friends and invite them to sign up to spread the knowledge and stay ahead.