AI Image Generation Has Grown Up

Apr 22, 2025

Imagine: you are producing an executive primer about how AI streamlines Accounts Receivable, and you need a professional grade image to pair with it. To invoke a sense of old, slow technology, you decide on the concept of a telephone handset and mailbox, standing together and waiting. Your design team works on a series of options.

Which would you choose?

Here’s the twist: all six were generated by AI. No stock sites. No design team. Just detailed prompts and the improving capabilities of image generation tools.

Image Generation Has Leveled Up

Over the past 12 months, AI image generation has crossed a threshold from "interesting toy" to "legitimate business tool." While the hype of the month has been Ghibli-style animation, the larger takeaway for businesses is that these tools can now create consistently professional, on-brand visuals that previously required dedicated design resources. You can now generate crisp line illustrations, minimalist diagrams, product mockups, and brand-aligned visuals.

For marketers, content producers and leaders, this is a paradigm shift – high-quality visual content is no longer bottlenecked by cost, time, or headcount.

The Test: Creating Images for Cerulean’s Knowledge Base

Cerulean needed to produce illustrations for our Knowledge Base, which contains a growing catalog of AI use cases. We wanted clean, minimalist visuals to distill complex concepts while maintaining a professional, high quality aesthetic. We tested six leading image generation providers: ChatGPT-4o (fka Dall-E), Leonardo AI, Adobe Firefly, Midjourney, Stable Diffusion 3, and Krea.

We asked each tool to create line art illustrations for each of these concepts:

1. AI in Accounts Receivable: A telephone handset and mailbox, conveying a sense of waiting.
2. AI for Invoice Reconciliation: A stack of marked-up invoices, conveying meticulous correction.
3. AI Sales Assistants: A person holding a notebook, conveying precision and purpose.

Here are the images generated by each tool, using the same prompts for each:

Assessment Framework: How We Evaluated These Tools

We evaluated each set of tools on the following criteria:

1. Accuracy: Did the tool follow detailed style and layout instructions?
2. Professional: Did the visuals feel brand-safe and usable for work?
3. Iteration: Could we refine outputs in real time without starting over?
4. Ease of Use: Could a non-designer produce high-quality results quickly?

🏆 Our Winner: Krea

Krea emerged as our top choice, with Leonardo as a close second. What specifically set Krea apart:

• Accuracy: High prompt-to-result accuracy without a steep learning curve
• Precision Control: It excelled at ultra-thin lines and technical schematics
• Real-Time Editing: Its tools let us refine outputs iteratively

Leonardo produced excellent results on single prompts but lacks Krea's iterative capabilities. Both outperformed more well-known tools like Midjourney and Adobe Firefly for our specific business needs. Check out the appendix for a detailed tool-by-tool assessment.

Mini-Tutorial: How to Write Image Prompts That Work

The quality of AI-generated images directly correlates with the quality of your prompts. Based on our testing, here are the key principles:

• Be Specific and Detailed: Include precise descriptions of what you want to see. Vague prompts produce vague results. You can also use negative prompts to specify unwanted elements to exclude.
• Use Artistic References: Mention specific art styles or techniques (e.g., "minimalist outline style," "architectural blueprint precision") to guide the aesthetic.
• Structure Your Prompts: Organize prompts into clear sections: subject matter, style elements, color scheme, composition, and mood.
• Control Composition: Specify positioning and viewpoint explicitly. Don't assume the AI knows where you want things placed.

Here's an example of a prompt that worked well:

A stack of marked-up documents with errors, line art illustration in a minimalist outline style, extremely thin and precise line work, technical schematic appearance, no shading or fill, ultra-fine continuous brown lines on a solid navy-blue background, objects shown in elevation view.

Tech Corner: How Prompting for Images Differs from Text

If you’re a ChatGPT/Claude/Copilot/Grok user, you probably already understand the basics of prompting (need help? Download our prompting guide). But image generation works differently under the hood, so your prompting approach needs to adapt.

What’s the same:

• Structured thinking works. Just like with text, the best image prompts break down the request into clear parts: subject, style, layout, tone.
• Iteration matters. Small tweaks to wording — or adding/removing constraints — can significantly change the outcome.

What’s different:

• Spatial and stylistic detail matters more. Text models infer context; image models don’t. You need to spell out positioning, scale, composition, and visual style explicitly.
• They start with noise. Most image models begin with a random grid of pixels and gradually refine it — which means vague prompts produce vague results.
• Fewer assumptions. Language models can “fill in the blanks.” Image models need tighter instructions to deliver something usable.

The takeaway? Be just as structured as you are with text prompts — but much more visually specific. This is why many business users give up after their first attempts. They're not being detailed enough.

We Still Need Human Designers

AI image generation isn't a replacement for human design expertise. While our top tools excelled at executing well-defined visual concepts, they don’t offer original creative direction and nuanced brand storytelling. The most effective organizations are finding a hybrid approach: using AI for rapid execution and production, while deploying human designers on higher-level creative strategy and art direction. This division allows designers to spend more time on creative thinking rather than repetitive production tasks—ultimately delivering more value while working alongside AI tools.

Key takeaways

AI image generation has reached the point where it can execute many design tasks. This isn't just about cost savings – it's about speed, flexibility, and the ability to iterate rapidly.

The business implications are clear:

• Early Adopters Leap Ahead: Companies integrating these tools now gain higher margins and a head start in developing AI management skills.
• Speed Matters: Image generation in minutes versus days.
• Consistency at Scale: Create hundreds of on-brand visuals without variation in quality.

In our November newsletter, we shared how global fashion retailer Mango uses AI-generated ad creatives. The result? Lower production costs allowing for more images tailored to specific segments.

What To Do Now

1. Try a tool like Krea or Leonardo for your next image task.
2. Practice effective prompting — it’s a learnable skill.
3. Ask your team about their go-to tools - there’s so much low-hanging upside from improving this workflow.

Questions about how to implement AI image generation in your workflow? Drop us a line – we're happy to share more details from our testing.

Appendix: Tool-by-tool assessment summary

Krea What We Liked: Excellent prompt accuracy, intuitive interface, iterative refinement.

What We Didn't: Requires deeper artistic and technical knowledge for handling finer compositional details.

Leonardo AI What We Liked: Excellent prompt-to-result accuracy, professional aesthetic.

What We Didn't: Limited iteration options, higher pricing tier required for best results.

ChatGPT 4o What We Liked: Direct integration with editing tools, easy to prompt.

What We Didn't: Less stylistic control, variable quality.

Midjourney What We Liked: Superior style control, unique aesthetic, high quality.

What We Didn't: Limited editing capabilities, Discord-only interface.

Stable Diffusion 3 (via OpenArt) What We Liked: Open-source flexibility, customizable models.

What We Didn't: Requires technical setup, quality varies.