Imagine... [Horizon 01 - Issue #9]

In partnership with

March 24 - March 31, 2025

In the last year, AI models have conquered writing, coding, spoken audio, and even music. Photography and image generation, however, proved to be much more difficult.

Yes, the progress was visible: if you remember, the first models generated distorted faces, hands with missing fingers, and shots with impossible physics. Things got better but it was all very gradual, without a “holy crap” breakthrough moment.

Until now.

The new ChatGPT-4o update is not simply better — it is jaw-droppingly superior. Overnight, we went from “this is a fun experiment” to “we will never purchase another stock photo again” or even “we will not do a photoshoot ever again”. From “how many more Canva licences do we need” to “do we need this many Canva licences for real, though.”

And ChatGPT is not even the only game in town. Reve dropped a model that could have been the bombshell of the week if OpenAI had not ruined their moment. Even more exciting, MidJourney is launching its new model later this week.

Add to this the news about H&M replacing (ok, “augmenting”, but come on, we all know the end game here) their fashion models with AI twins. These are unprecedented times for the creative industry as a whole. It is a privilege to be your guide.

Peter & Torsten

Cold Email Setup Offer

We started sending 10,000 cold emails per day, and scaled a brand new B2B offer to $108k MRR in 90 days. Now, you can have the same system set up (completely done-for-you) inside your own business - WITHOUT going to spam, spending thousands of dollars, or any manual input. Close your next 20 clients easily. We’ll set up the tech, write your scripts, give you the leads, give you the inboxes, and the sending tool - all starting at $500/mo.

Get Pricing & Details

ChatGPT’s new image-generation features: from last to first in one update

Image generation was not, traditionally, one of OpenAI’s fortes. The underlying model, Dall-E, just wasn’t great or even just good. With the latest multimodal update of 4o, however, ChatGPT catapulted from last to first.

Unless you’ve been on a retreat with no internet this past week, you have surely seen Studio Ghibli-style images flooding everything. Memes, family photos, historical images, nothing got spared:

And it’s not just cartoons. The new model is just incredible: it handles text in images much better than before, and complex prompts with many elements reliably come out with correct details. For example, it can correctly bind attributes for 15–20 distinct objects in one scene without mixing them up, a big improvement in accuracy over previous image generators.

The end of stock photography

This level of prompt adherence makes stock photography all but obsolete (especially when used online — print is a different beast because of resolution requirements). You will never need to search for another photo, ever. You just generate the one you want.

Let’s see an example prompt:

Generate a photorealistic image of a thin but muscular black male in her late 20s training with dumbbells in a gym. He is doing biceps curls with 40 pound dumbbells. He is wearing blue short pants and a white Nike t-shirt. The lighting is heroic, with darker corners and a spotlight on the man, but the gym setting is visible with machines and free weights on the sides. He is struggling with the training set but determined, ready to conquer the weight. He is exhausted, sweating, but winning. The image should convey heroism. This is a professional ad photo, but it's very realistic.

And the result:

This image is, basically, perfect. The logo placement, the anatomy, and even the 40-pound markings on the dumbbells. But we can make it even better:

Great now can you please put this logo attached on the top left corner? But it is in black, you need to convert it to white first (just invert the colors really). And add the text "YOU ARE WINNING" in white with a sans serif font to the bottom, aligned to the center.

Note that the logo I’ve attached to the prompt is black on a transparent background, and we need it in white to work with this photo. No problem, says ChatGPT, and executes the task just as instructed:

Now, if your first reaction is, “But it’s not the same guy,” you are right. He became even more heroic, and I would argue that the composition is even better than the original. You could put this creative on social media just as it is, with no editing required.

Sure, Photoshop might be safe for now. But is Canva? Just look at this:

Please generate an image for me. It should be coupon I can upload to Facebook. The color of the coupon should be #F70006 as that is our brand color, and the text should be white. It should say with big letters: 20% OFF YOUR NEXT PURCHASE. In smaller text under it: Valid until April 30, 2025. Under it: Applied automatically, no coupon code required.

The result is basic but usable. No image model could ever produce anything like this due to limitations with text rendering:

Now, let’s make it a bit fancier:

Make it more looking like a coupon, with a tear-off part and some squiggly lines

No, Canva is definitely not safe:

Let’s push the limits of creativity and ask ChatGPT to come up with a funny comic strip:

Generate a 4 image comic about AI trying to take people's jobs but constantly failing in a funny way.

And the result:

This is genuinely funny and perfectly executed, with the callout bubbles and text elements.

I am sure you’ve noticed that no marketer appears in this comic strip. Use this information as you wish.

What could have been: Reve Image 1.0

A Palo Alto startup, Reve AI, has launched Reve Image 1.0, an advanced text-to-image model focusing on prompt accuracy, aesthetics, and even typography.

Sounds familiar?

Reve is wonderful: creators can generate images from scratch and modify existing images, like changing a product photo’s background or a design’s color scheme, just by typing the request. It even supports reference image uploads to guide style, helping match a specific brand look or inspiration.

Sounds familiar?

One standout capability is text rendering. Reve Image 1.0 produces clear, legible text (think logos, posters, or social ads with crisp fonts). It also excels at multi-character scenes, reliably positioning and differentiating several people or objects where older models might fumble.

Again, sounds familiar?

Reve could have been the drop of the week, but ChatGPT stole the show.

Why should you still consider Reve?

In our tests, Reve performed marginally worse than ChatGPT. Look at the same poster with the same prompt:

It looks decent enough, but it is not very heroic, right?

But using Reve is sometimes just more convenient. First, you can generate multiple images from a single prompt, giving you more flexibility and faster iterations. Second, Reve does adheres to aspect ratio instructions, while ChatGPT tends to ignore them.

Reve might not produce the very best results, but it is a capable, reliable fallback option when ChatGPT can’t produce what you imagined.

MidJourney 7 drops this week

Funnily enough, ChatGPT’s reign might be very short lived indeed. The popular image generator MidJourney is rolling out its highly anticipated Version 7 model this week, and it is expected to deliver a major leap in image quality and generation speed, coupled with a more intuitive user interface.

Under the hood, MidJourney V7 is being rebuilt from the ground up – the team has hinted at a new AI architecture and fresh training data for sharper, more coherent images than ever. Essentially, V7 should produce visuals that are not only prettier and more photorealistic but also better understand the nuances of user prompts to give you exactly what you envision.

MidJourney has always flown somewhat under the radar, thanks to its peculiar Discord-first strategy, but it’s been the undisputed king of image quality before the ChatGPT update. We expect v7 to regain that spot or at least put up a good fight.

H&M’s AI-Generated Model “Twins” Spark Debate

Global fashion retailer H&M is stepping into AI by creating digital “AI twins” of 30 human models for use in its marketing and social media campaigns. In collaboration with a tech partner, H&M generates virtual replicas of these models to showcase clothing and accessories.

The company is taking a human-centric approach – meaning the real models have given consent, will retain rights over their digital likeness, and will be fairly compensated for any use of their AI-generated doubles. In practice, you might scroll Instagram and see what looks like a real H&M model in an outfit but notice a small watermark or label indicating an AI-generated image. H&M says it will clearly mark these images as AI-made to maintain transparency with the audience.

The move shows how AI can scale up content production in fashion marketing – digital models don’t need rest and can wear any outfit in any setting on demand. For H&M, this could mean a faster turnaround for showing new collections online and the ability to try more styles or creative concepts without organizing full photoshoots every time. It also opens up the possibility for the brand (and even its partners or competitors down the line) to license these virtual models for other uses since the digital “twins” are assets that can appear anywhere with a few clicks.

However, as you can imagine, H&M’s AI modeling experiment has sparked debate across the creative industry. If digital replicas can do the job, will there be fewer gigs for real models and the crew (photographers, makeup artists, etc.)? H&M has preemptively worked with model representatives (like the Equity union in the UK) to ensure protections; for example, models can veto the use of their AI twin and are paid similarly to they would be for a photo shoot.

The broader concern remains that as this practice spreads, the role of human creativity and labor in fashion could diminish. It’s a story to watch, as it raises important questions about authenticity, jobs, and the future of creative work – much like how CGI impacted the film stunt industry, AI could transform modeling. For now, H&M is positioning it as a tech-forward complement to, not a replacement for, traditional imagery.

ByteDance’s InfiniteYou – personalized portraits at scale

Talking about digital replicas…

TikTok’s parent company ByteDance just introduced a new tool called InfiniteYou, which focuses on photorealistic portraits. Unlike conventional AI image models that often struggle to maintain a person’s likeness across different images, InfiniteYou is designed to keep the identity consistent while changing contexts, styles, or poses. In practical terms, if you have a headshot of someone, InfiniteYou can produce numerous new photos of that person – in different outfits, backgrounds, or artistic styles – and still look like the same person each time.

The immediate use case

InfiniteYou is a huge opportunity for marketers dealing with branded imagery or personas. For instance, imagine a retail brand’s ambassador or a model – InfiniteYou could generate dozens of themed photos of that same model for different campaigns (summer, winter, sports, formal) without additional photoshoots. A DIY version of the H&M story, if you will.

Customer success teams, on the other hand, might use it to personalize customer education materials – e.g., creating friendly tutorial characters that remain familiar across all illustrations and guides. On the flip side, this tool raises the bar for realistic deepfakes in the creative space, so companies using it will need clear permissions and ethical guidelines. But used legitimately, InfiniteYou offers a tantalizing ability to re-imagine one person in endless ways – a powerful asset in any creative toolkit.

Gemini 2.5 needs marketers. Do marketers need Gemini 2.5?

Google has unveiled Gemini 2.5, billing it as its “most intelligent AI model” yet, with built-in thinking capabilities for advanced reasoning. In fewer buzzwords: this latest Google model can internally plan out multi-step solutions before answering, which helps it handle complex questions and tasks more accurately. It’s a multimodal model, so it isn’t limited to text – it can also work with images or other data forms, much like OpenAI’s GPT-4.

And it is very good, leading the LMArena leaderboard by a large margin:

(Yes, the updated ChatGPT-4.0 that introduced the new image generation is better than ChatGPT-4.5. No, OpenAI’s product strategy does not make sense.)

Gemini 2.5 writes better copy than ChatGPT and writes better code than Claude, and it’s the most powerful, best model on the market now. You can try it yourself if you are on the paid tier of Gemini or for free in Google AI Studio.

That said, Google might need to work on marketing Gemini 2.5 itself. Gemini’s capabilities are impressive on paper, but Google’s challenge is persuading users that it meaningfully outperforms existing solutions for day-to-day tasks.

What is the USP? Where is the brand? Why is Google so bad at marketing the best product available? If you are from Google, drop us a message—we’d love to chat.

Until next week — Peter and Torsten

Quick notes of the week

* Sponsored

• We upgraded our Beehiiv account, allowing us to launch a much better web experience for our readers. Please look at our new site and let us know what you think.

• Speaking of Beehiiv, if you want to launch your newsletter, use our link and get a 30-day trial + 20% OFF for 3 months.*

• Not everything is image - text is still important, especially if you need to write high-quality SEO-focused content with AI. Many tools are great for that. We particularly like AISEO for its single-focused use case.* 

• Most marketers use ChatGPT, the most popular tool available. Yet, when we ask people, “Have you designed a Custom GPT for a specific use case?” They shake their heads and say back, “I don’t know how.” This is where CustomGPT AI is helpful. Give it a try.*