1. Introduction: The Myth of ChatGPT’s “Omnipotence”

In the world of visual design, it’s getting harder to escape one pressing question: has AI already “done it” better than us? ChatGPT-4o, the latest tool in OpenAI’s arsenal, can now generate not only text but also images – and at a staggering speed. You type in a prompt, wait a few seconds – and there it is. Scene, model, lighting, product. Done. Clean. Polished.

For many – it’s a miracle. For a designer – it’s... ambivalence. On one hand: precision, control, repeatability. On the other: unease.

Something feels off. The image is very correct. The composition works, but it’s not memorable. The lighting is accurate, but... it lacks that unique “twist.” Over time, you start to realize these images aren’t so much created as they are assembled from pre-made answers. As if the model had its own aesthetic “pattern book” – and kept pulling from the same safe palette. That’s no accident. It’s calculation.

And this is where the real conversation about AI aesthetics begins: not about what the technology can do, but what it means for imagery. For expression. For visual culture.

2. Homogenization and Visual “Flatness”

2.1. Template Thinking

Dominant Pattern: ChatGPT-4o has been optimized for accuracy and consistency – which is exactly what makes it so predictable. The models are trained on datasets full of commercial imagery, resulting in visuals aligned with stock aesthetics: smooth, clean, stripped of ambiguity.

Apparent Quality vs. Depth: The images are sharp, saturated, often even perfect – but lack atmosphere. Perspective tends to be flat, compositions almost exclusively frontal. The frames resemble catalog illustrations – correct, but lifeless.

2.2. Narrow Plastic Exploration

Flat Light and Shadow: Many generated images look like they were shot on an overcast day – diffuse lighting, no dramatic shadows. Achieving the effect of sharp, high-contrast shadows is difficult.

Uniform Textures: Skin, fabric, background – everything often looks similarly sterile. And therefore… similarly banal.

Lack of Expressiveness: The models lack compositional intuition – even when prompts are “amped up,” it’s difficult to achieve an outcome that deviates from what is canonically correct. All elements compete for attention.

3. Contrast: Why Midjourney Still Inspires

3.0. Comparative Analysis Based on Personal Testing

While preparing this essay, I conducted a series of comparative tests using identical prompts and input materials. I directly compared images generated by ChatGPT-4o with those from Midjourney. The results of this analysis were crucial in forming most of the observations and conclusions in this essay.

My micro-research findings were unambiguous:

ChatGPT-4o exhibits high consistency and precision in detail rendering, but the images often feel too clean, too literal, and formally safe.

Midjourney produces stylistically more diverse images, with richer light-shadow dynamics and stronger moods. Even with the same input material, the results varied significantly – in a good way – offering room for visual interpretation and inspiration.

That said, Midjourney has developed its own visual clichés (e.g., characteristic lighting or hyper-detailed textures), recognizable to the trained eye. Despite the claimed diversity, even its images can fall into stylistic repetition – though it offers more opportunities to break these patterns, especially with the growing availability of personalization and the tool learning the user’s aesthetic style.

Midjourney’s added value lies in better control over variation – through parameters that lead to effects beyond the standard. This makes it particularly suited for conceptual design phases, where ambiguity and visual boldness are essential.

Fig. 1. Comparison of ChatGPT-4o and Midjourney Models (using the same prompt and custom input materials):

3.1. The “Crazy” Flow

Midjourney embraces different values: not precision of reproduction, but strong visual impact. The images generated are full of contrast, rich in form, often painterly, fantastical, or even grotesque. This expressive power is one of Midjourney’s most frequently praised features by users.

Importantly, the tool doesn’t just execute prompts – it often interprets them artistically, giving results a unique character. This opens a new space for curatorial and art-directorial work. It’s not about accurately reproducing reality or physical details, but about creating new, unexpected visual worlds. Midjourney thus becomes a tool for creating visual narratives, not just illustrating them – crucial for concept-driven work, where what something feels like matters just as much as what it looks like.

In contrast, current ChatGPT models struggle with abstract concepts beyond surface-level meaning. The images, though “pretty,” often remain shallow and repetitive in form and expression.

Midjourney, known from the outset for its surreal, experimental aesthetic, serves as a counterweight to the predictability of models like ChatGPT-4o. Its stylistic boldness and visual risk-taking are highly appealing to creators – but that very popularity is leading to the rise of new, recognizable “AI-generated aesthetics.” This vivid style, while initially attractive, is now frequently overused, forming another kind of visual cliché. Even the most creative tools can contribute to homogenization – unless users consciously challenge the default.

3.2. Remixing and Unpredictability

Parameters like --chaos allow the deliberate introduction of creative randomness. Midjourney often interprets prompts in unexpected ways, adding stylized or surprising elements. Users, too, can disrupt its perception with varied inputs or fine-tuned parameters. In this way, it moves closer to the role of co-creator, not just executor.

3.3. Comparative Summary

4. Why Is There “Flatness” in ChatGPT-4o’s Images?

Although models like ChatGPT-4o remain partially opaque (no full access to training data or weight mechanisms), working with them daily and analyzing available examples and expert insights allows us to identify some likely aesthetic biases. Observations suggest the system may have been deliberately tuned to ensure maximum predictability, cleanliness, and alignment with prevailing visual norms. As a result, its images resemble idealized catalog visuals more than expressive artistic visions. This is logical from a commercial application standpoint – but for many creators, it means a loss of space for authenticity and experimentation.

• Training Data Bias: Commercial and stock imagery shape AI aesthetics toward the “universal” and predictable.

• Client Optimization: Models are built to meet business expectations: images must be “clean,” “pretty,” “correct.”

• Low Randomness: ChatGPT-4o minimizes variance, reducing errors... but also creative serendipity.

Fig. 2. Comparison of ChatGPT-4o and Midjourney Models (using the same prompt and custom input materials):

5. The Rare Becomes Even Rarer

As more creators turn to the same models, visual monotony accelerates – not as a matter of taste, but as a systemic phenomenon. The images are correctly good, but as if from the same gallery – differing in subject, but not in style, since algorithms operate within the same statistically optimal patterns. Truly exceptional things – those that build new aesthetic or conceptual quality – are now even more unique and harder to achieve. AI can’t generate them without a nudge from “analog” intuition – human error, imperfection, deliberate risk. That’s the paradox: what used to be unwanted noise in the creative process is now the most desirable ingredient of originality.

6. Breaking Out of Flatness

• Experiment with hybrids: Combine AI with photography, drawing, scans, parameters.

• Change parameters and contexts: Force AI to “make mistakes” – that’s often where inspiration hides.

• Use personalized models: Fine-tuning your own style is the path to uniqueness.

• Mix tools: Don’t stick to one AI – treat them like different brushes in your palette.

• Hack the tools: Treat AI not as a closed system but as raw material for transformation. Don’t settle for ready-made outputs – question their logic, remix techniques, combine generators in unexpected ways. Work across tools: if one is too predictable, use its output as a starting point for another process. Look for the cracks – the unintended spaces where your authorship can emerge. That’s where true design begins.

7. Conclusion: Stability as a Creative Ball and Chain

ChatGPT-4o is an excellent tool… for precise applications. Its greatest strength – stability and consistency – becomes a trap in creative work. Where there’s no variability, there’s no room for chance – no crack through which inspiration might slip. And it’s precisely these cracks – imperfect, surprising, unintentional – that form the foundation of every true creative act. Still, the contrast between “precision and predictability” and “unpredictable expression” is stark.

It’s also worth noting that the immense hype and widespread use of image generation within ChatGPT’s interface may have long-term effects on collective visual literacy. As this tool becomes mainstream, more creators and viewers start seeing AI-generated imagery as a new aesthetic standard. This could further flatten visual culture – not because the tool is bad, but because it too uniformly shapes viewers’ imaginative range. When what’s generated becomes what’s acceptable, we risk blurring the line between aesthetics and aestheticization – between image as message and image as decoration.

Midjourney isn’t better – it’s different. But in my view, especially for messy, creative conceptual work, it offers more space for breaking patterns. Where ChatGPT-4o provides stability and consistency, Midjourney enables visual risk, intuitive prompt-breaking, and the emergence of unexpected effects – which often become turning points in a project. And that’s why it remains a haven for those who, in the generative age, seek not just images, but voice. Something personal. Something that still resists being fully predicted. And that, despite AI, remains profoundly human.