How AI Image Generation has Improved in 1 Year

This week, I thought it would be fun to take a trip down memory lane at the images generated in previous Matrix issues. 

Shoutout to Dina Otero, VP of Demand Gen at Mission for the idea.

A Prompt Review

When I started the Matrix in early 2024, AI image generation was already impressive, but it certainly had its quirks. 

Models would sometimes struggle with basic physics, text, character consistency, or understanding complex scene compositions. 

Fast forward to today, and though image generation still struggles, these models are performing a lot better. 

I took a few prompts from our early newsletters and ran them through today's AI systems. The side-by-side comparisons are very interesting.

February 2024

PROMPT: “A picture of Cookie Monster in a cookie house that he wants to eat but can't”

What's Changed

If you’ve been here from the beginning, you might remember that this picture on the left was the image that started it all! I remember being impressed with the output, but compared to today’s outputs, it feels a bit more elementary.

The 2025 image shows dramatic improvements in photorealism and detail rendering. You can tell that the texture work is more sophisticated with individual fur strands with realistic lighting and shadows, while the 2024 version has a more synthetic, "plastic-y" appearance. 

The gingerbread house in the newer image shows a much better understanding of materials, with proper icing drips, realistic cookie textures, and accurate lighting that looks like something I actually would want to eat.

Behind these visual improvements are technical shifts. 

AI models moved from traditional diffusion architectures to transformer-based systems like those powering GPT-4o's image generation.

Training increasingly relies on synthetic data and flow matching techniques that create more efficient and higher-quality outputs. 

The result is imagery that users report as indistinguishable from professional photography. Equally amazing and scary, in my opinion.

March 2024

PROMPT: “Create a photorealistic image of Creator AkiraToriyama and Son Goku from Dragonball Z designing a groundbreaking new AI product in a modern Silicon Valley workspace. They should be engaged in an intense discussion, with holographic product renderings and AI code surrounding them.”

While I was running these comparison tests, I hit something that tells an even more interesting story about AI development.

One of our old prompts featured a popular anime character. In early 2024, most AI systems would generate these without hesitation. 

But today, complete rejection due to content policy violations.

This is responsible AI development.

The Maturation of AI Governance

The AI industry is growing up. 

Companies have implemented sophisticated content detection systems that can identify copyrighted characters, potentially problematic content, and intellectual property concerns. 

AI companies understand copyright, fair use, and intellectual property law much better than they did in the Wild West days of early 2024.

As these tools moved from experimental to mainstream, companies shifted from "move fast and break things" to "move thoughtfully and build sustainably."

The guardrails we see today represent months of careful consideration (and lawsuit headlines) about how these powerful tools should and shouldn't be used.

April 2024

PROMPT: “Create an image of Scottie Scheffler celebrating his win at the Masters in Augusta, Georgia. Scottie is holding a trophy. Scottie is wearing an orange polo, a white ball cap, and blue pants. He is surrounded by cheering fans. It is a gorgeous, sunny day. TV crews are filming the celebration.”

Now this one is just downright wild to me. The improvements in facial accuracy and likeness generation shows us how much AI models have evolved to handle real people. 

2024’s image shows a generic golfer celebrating. If you saw that image at the time, you could maybe assume it's Scottie Schleffer, since he had just won the Masters.

Fast forward to today’s image and WOW. 

You can tell AI now understands facial structure, proportions, and distinctive features due to enhanced training datasets and improved facial recognition capabilities. 

Modern models can now process prompts containing real people's names and generate more accurate representations rather than just generic faces. It even captured his tan lines!

Outside of actually looking like Scottie, the 2025 image also shows that AI has advanced in overall scene composition. 

Better crowd rendering, more realistic golf course details, and improved lighting make this feel like a legitimate picture (just don’t zoom in and look at the hands!).

What This Means for You

Image generation’s evolution has three major implications for how you should think about AI in your organization:

1. Quality Expectations Should Rise

If you implemented AI image generation in early 2024 and were disappointed with the results, it's time for a second look. Tools that you thought had promise in 2024, but weren’t quite there, are worth another try.

2. Compliance is Built-In…

…But still do your due diligence.

The stronger guardrails mean you can deploy these tools with greater confidence that they won't create content that lands you in legal hot water. But remember, always keep human-in-the-loop.

3. Innovation Pace is Accelerating

If this much has changed in 18 months, imagine where we'll be by 2026. The companies that start experimenting and building expertise now will have a massive advantage.

The Framework That Still Matters

While the technology has improved dramatically, the fundamental framework for successful AI implementation hasn't changed. You still need to:

1. Define Clear Expectations: Know exactly what output you're hoping to achieve
2. Design Your Workflow: Think through your creative process as a "prompt outline"
3. Test, Analyze, Iterate: This remains the path to success with AI

The difference is that today's tools will reward your efforts with way better results.

The Bottom Line

Whether you're creating marketing materials, developing product prototypes, or building entirely new creative workflows, today's tools are leagues above what we had just 18 months ago.

If you enjoyed this week’s Matrix, let me know. I had a lot of fun writing this and would love to do another issue comparing more images from previous Matrix issues with the capabilities of models today.

Until next time,
Ryan

Now, since today’s issue was full of AI-generated images, I thought it would be nice to see some real images from my vacation over the last two weeks. I went to Legoland in Billund, Denmark, then some pics from my travels around:

• Riga, Latvia 
• Tallinn, Estonia
• Gdansk, Poland
• Oslo, Norway
• Lubeck, Denmark