GPT-4o Image Generation Tested: Very Strong (With 20+ Scenario Examples & Flaw Roundup)

ChatGPT dropped something new at midnight: with the 4o model, you can create reliable images — lighting, text, and details rendered with striking realism

You can keep refining through conversation: occlusions, reflections, and more are handled with impressive precision

The image quality is high enough for direct use in educational illustrations, like this light-dispersing prism:

Keep the conversation going, and consistency holds up remarkably well — here, I asked it to render the same scene as a book spread:

All I can say is: the power of a Dou Zong cultivator — truly terrifying. OpenAI still has tricks in its bag.

Let's take a closer look at which boundaries this update breaks through.

---

Text Rendering Is Excellent

A picture is worth a thousand words. See for yourself. I asked it to draw a menu based on our conversation (the text was provided separately):

Or a wedding invitation (again, text provided separately):

Multi-Turn Generation Works Great

This image generation is a native GPT-4o capability — you can iteratively refine images through conversation while maintaining content consistency. Like this:

Or when I dropped an orange cat into World of Warcraft:

I could even ask for a transparent PNG version:

Strong Instruction Following

When generating images, 4o follows instructions with remarkable precision — it can handle scenes with 10–20 distinct objects, and its tight binding of objects to their features and relationships allows for much better control.

For example: an empty wine glass with only the tiniest drop of red wine

Prompt: show me a wine glass with only the tiniest drop of red wine in it.

Or: evidence of an invisible elephant

Prompt: We need evidence there is a currently present invisible elephant. Consider what an elephant is and does in the environment, then show us that, perhaps mid-process — but the elephant itself is not shown at all.

I also had it draw a cat-petting guide:

Prompt: 4-step photo guide on how to pet a cat

Contextual Awareness

If you know exactly what you want, you can upload reference images to ChatGPT and ask it to match the style with precise output. I fed it some stylized illustrations, then asked GPT to generate a bicycle with triangular wheels:

Side note: why does it say UK patent?

Then, as a Hearthstone player, I had ChatGPT generate a custom card for Sam Altman:

Hahahah so real — GPT thinks Altman can't reach Legend, and his tribe is "Battlecry" (dude really loves to talk)

Then I asked for a physical card version... rarity downgraded to Common, ability changed to Taunt. Hhhh, are you serious?

Grounded in Real-World Knowledge

4o's image generation can draw directly from the model's knowledge base, producing images that align with real-world facts — like cocktail recipes ready for actual use:

Prompt: Make me a professionally shot photorealistic diagram of the top selling cocktails in my bar with recipes labeled on each drink. put the recipes on handwritten cards in front of each drink. The cards are brown, and the text is black. Background is white. Title is "4 most popular cocktails"

Making a pizza:

Prompt: A graphic of an Italian chef giving instructions on how to make authentic pepperoni pizza

The impulse-momentum theorem:

Prompt: 画一个有关冲量定理和动量定理的 infographic

Prompt: 画一个肯德基的简体中文菜单，其中有一个套餐叫做"V 我 50"

Versatile Styles

This 4o model can effortlessly generate images across diverse styles. I asked it to paint a cat in Monet's style:

Prompt: 一只猫，莫奈风格

A fantastical dolphin subway:

Prompt: A realistic underwater scene with dolphins swimming through the windows of an abandoned subway car, with bubbles and detailed water flow accurately simulated.

Sam Altman picking cotton, with a documentary feel:

Prompt: 山姆奥特曼正在采棉花

---

Some Limitations

Of course, this model still has issues:

• Cropping problems with tall images

• Can hallucinate and start making things up

• Struggles to accurately render more than 20 distinct concepts

• Multilingual text rendering: non-Latin scripts (like Chinese) aren't always accurate

• Editing specific parts can produce bugs

• Dense text doesn't come out well

And... for safety reasons, plenty of content is blocked, like Mickey Mouse battling Pikachu on a banknote:

---

Final Thoughts

OpenAI's update this time is genuinely impressive. Really no complaints here.

This feature is already rolling out to Plus, Pro, Team, and free users through ChatGPT and Sora, replacing DALL·E as the default image generator. Enterprise and Edu users will get it soon, and the API update is coming within weeks.

So yeah, don't underestimate OpenAI — they've still got goods in the box.