How we use AI for marketing

This post summarizes our experience using AI tools for generating content for marketing campaigns.

To set the record straight - we are not interested in creating anime girls nor famous actors eating spaghetti nor astronauts riding the horses. We use AI for commercial purposes, i.e. helping the clients to sell more stuff.

Are we there yet?

Can AI be used for marketing imagery and videos?
Are we there yet?
TL;DR: Almost!

Skip to the examples if you don't feel like reading the text.

When we first heard about Stable Diffusion, we (like many others) immediately saw the marketing potential for image generation. It seemed that unlimited imagination was finally within reach. In practice, however, a few steps are still needed before AI becomes truly great for image and video production.

Sometimes, AI capabilities feel like a magic; other times they are hit-or-miss. In the majority of cases, post-production work is still required.

Image generation

Image generation is tricky. Text-to-image has evolved the most of all visual AI modalities. Many models can create striking imagery when the prompt is spot-on, but when it comes to product photography - crucial for commercial work - text-to-image alone isn't sufficient. Image-to-image techniques become necessary.

Customers need real product - be it a car, house, yarn skein, or anything else they sell - to appear convincingly in the image. The biggest obstacles here are consistency and speed of generation.

Anyone who tried to convince a model into using a specific product and keeping it identical accross frames - maintaining shape, color, label, texture, etc. - knows that it rarely works without several attempts. You can't rely on a prompt alone. LoRAs, ControlNets, and similar helpers are needed to steer the AI toward your vision. Tools such as ComfyUI let you build sophisticated visual workflows, while developers can script the process in Python if they prefer code over node-based interfaces.

Even with those aids, our experience shows that the generated images of a product and its branding are not always consistent, which remains a major hurdle for commercial use.

Video generation

With video, the need for improvement is even more evident.

Consider a recent AI generated ad campaign created for a large, deep-pocketed company. Because budget wasn't a constraint, the project could afford top tier talent, state-of-the-art tools, and the latest models.

Nevertheless, a close inspection of the final output reveals significant shortcomings. Details that should be seamless appear off, and even seasoned viewers of AI generated footage can sense that something isn't quite right. Often subtly, but unmistakably.

That said, the effort is worth appreciating. It demonstrates what is technically possible today while also highlighting the many promises that lie ahead.

Remember, we are only at the beginning. It's therefore crucial that individuals and companies experiment with these new technologies, because widespread use is the fastest path to refinement and improvement.

Coca-Cola Christmas video ad (2025)

The ad was produced by five AI specialists who assembled more than 70 000 short clips over a period of roughly 30 days.

Most people scrolling through social media at “the speed of thought” probably won't notice that the piece was generated with AI. Or, if they do, they may not care. In that sense the result can be acceptable for the advertiser.

A close viewing, however, reveals a number of typical AI artifacts: slightly distorted logos, inconsistent truck license plates, odd movements of the people, and nonsensical lettering on signs.

These inconsistencies illustrate how difficult it still is to get AI to render fonts, letters, and words accurately. In general, the same challenges apply to hands, vehicles, animals, fruits, and virtually any other object that requires precise detail.

Compared with the 2024 AI generated Coke Christmas spot, the 2025 version looks noticeably sharper. The progress made in a single year is evident.

Takeaway: Generated images - and especially videos - require substantial post-production work. Never underestimate the time needed to clean up AI output.

Models

Nano Banana

We had experimented with image models before, but the arrival of Nano Banana Flash was a genuine breakthrough for us. It felt like magic: suddenly we had an image model that could generate decent visuals very quickly. Speed matters a lot in marketing because iterating on concepts must happen in near-real time; otherwise the creative spark can fade.

Qwen Image

Another highlight was Qwen Image (and its companion Qwen Image Edit). The results are strong, and the model's permissive license lets you use the outputs without worrying about restrictive terms.

Some commercial models are overly cautious. They refuse to generate images that might violate undocumented safety rules. In a marketing context, that rigidity can be a hindrance when a campaign needs creative freedom.

Z-Image Turbo

Our most recent addition is Z-Image Turbo. It delivers excellent quality, rapid generation, and runs with only 6 billion parameters. Some publications claim sub-second generation; in our tests on an NVIDIA A100 GPU we see roughly 3 seconds per image, which is still very fast and eliminates the long waits that older models imposed.

Grok Imagine

We employ Grok Imagine for video generation. It is relatively swift, though it has a few quirks. The native output format isn't exactly what our pipeline expects, but with our internal post-production tools we can convert the footage into a usable form.

Other models

The models listed above are not exhaustive. We continually evaluate and adopt additional systems as the landscape evolves. New releases appear frequently, and we plan to keep testing emerging options.

What this means for marketers

These are exciting times, but it's important to acknowledge that AI generated imagery is still in its early stages. For us, two primary obstacles remain:

• Consistency - keeping objects, colors, and branding stable across multiple renders.
• Speed - reducing generation latency enough to sustain rapid creative iteration.

Text generation

We were enthusiastic about text generation at first, but lately the results feel a little bland.

We turn to it when we need copy inspiration, yet we rarely can publish the generated text verbatim. Convincing the models to adopt a client's authentic voice has also been a challenge.

Our toolkit

We rely on a mix of AI models, scripting environments, and creative applications to move ideas from concept to finished asset. This isn't an exhaustive inventory - as new models and workflows emerge we'll adopt whatever helps us stay efficient.

• Various large language models (LLMs)
• ComfyUI
• Python
• Internal tools - simplified versions are available as free data apps for quick prototyping
• Affinity suite
• CapCut
• Kdenlive
• GarageBand
• macOS, iPadOS, Linux

Real examples

Below are a few assets we created for the campaigns run by Makaron.cz, a Czech e-shop that sells yarn for crocheting and knitting.

The niche is ideal for testing AI generated assets - yarns come in many thicknesses, textures, colors, and label designs, and every visual element must stay consistent across a campaign.

It's also honest to say that not every asset we considered “great” performed well. Fortunately, we're data obsessed and track every metric, so under-performing ads are switched off quickly and replaced with fresh iterations.