Review: OpenAI’s New Image Generator Is Great Again

OpenAI has just overtaken the AI image generation race once more.

The tech giant’s integration of native image generation directly into ChatGPT via its GPT-4o model is not an incremental change but a major overhaul of the model, vaulting it to the front of the class.

Within hours of its release yesterday, the model quickly went viral, with anime-style creations flooding social platforms and showcasing technical capabilities that leave DALL-E 3 in the dust.

The new model can easily compete against dedicated image-generation platforms while eliminating traditional workflow barriers.

The $20 monthly ChatGPT Plus subscription now delivers a comprehensive creative ecosystem that would previously require multiple specialized tools and subscriptions.

The visual showdown: GPT-4o vs. industry leaders

We compared the model against Flux (the best open source image generator) and Reve (the best closed source image generator), and here is what we found

Realism

Prompt: A high-resolution photograph of a bustling city street at night, neon signs illuminating the scene, people walking along the sidewalks, cars driving by, a street vendor selling hot dogs, reflections of lights on wet pavement, the overall style is hyper-realistic with attention to detail and lighting, a neon sign says “Decrypt.”

Our urban nightscape challenge—requiring sophisticated light physics, crowd rendering, and architectural precision—revealed distinct performance profiles across competitors.

ChatGPT delivered impressively vibrant environments with neon signage, creating rich reflections across meticulously rendered wet pavement.

While excelling in crowd dynamics and element inclusion, the minor perspective inconsistencies occasionally betrayed its synthetic nature.

The lighting was also goo,d but sometimes veered into theatrical rather than naturally urban. It also was not the best at reflections, but this is something that only the most picky ones would catch. It also generated legible neon signs besides the “Decrypt” one, which also adds to the realism.

Reve is for us the winner through good light physics modeling, particularly the subtle interactions between neon sources and reflective surfaces.

Its cinematic framing and atmospheric elements (steam wisps, motion blur) created superior dimensional authenticity. However, it reduced crowd density, which was a clever hack since it didn’t have to generate a lot of faces, making it harder to spot unrealistic details.

The system prioritized mood over literal prompt adherence.

Freepik Mystik (Flux) interpreted our prompts through a different lens and was the model that deviated the most from the realistic style.

It mixed Asian with Western lettering, generated different Decrypt signs instead of just one, and suffered from technical limitations in human rendering and dimensional depth.

Its reflective surfaces lacked the physical accuracy displayed by ChatGPT.

Winner: Reve narrowly secured the realism crown through superior rendering of complex lighting interactions. ChatGPT established itself as a remarkably close second, particularly impressive given its integration within a broader multimodal system rather than a specialized image generator.

Prompt adherence and spatial awareness

Prompt: A dog with a red hat standing on top of a TV showing the word ‘Decrypt is the best Crypto+AI media site in the world’ on the screen. On the left there is a blonde woman in a business suit holding a coin, on the right there is a robot standing on top of a first aid box, a green pyramid stands behind the box,. The overall scenery is surreal. A cat is standing upside down on top of a white soccer ball, next to the dog. An Astronaut from NASA holds a sign that reads “Emerge” and is placed next to the robot. Keep a widescreen format.

How intricate could instructions become before systems failed to render elements in their specified relationships?

This is what we wanted to test here, so realism, beauty, or other aspects were not as critical.

Current models are so good at prompt adherence that we need to tweak our testing prompts.

We progressively increased complexity in our prompt until reaching a surrealist composition requiring precise placement of over 25 distinct elements. All the other models failed in previous stages

ChatGPT demonstrated extraordinary prompt fidelity, accurately rendering 23 of 25 specified elements in their correct spatial relationships.

The achievement represents unprecedented prompt comprehension, like watching an experienced artist transform detailed verbal instructions into nearly perfect visual execution with only minor deviations.

For those picky enough, the only two major bugs we found were the cat not being upside down and the green color spilling from the pyramid to the first aid kit.

Freepik Mystik showed significant comprehension degradation, correctly rendering approximately half the requested elements while misinterpreting spatial relationships and modifying key components.

It was the model that failed the test first. The colors spilled to different elements of the composition (the red hat generated a red TV and a red wall), and the concepts also spilled—the dog on the TV spilled to generate an astronaut dog, for example.

Reve demonstrated poorer prompt fidelity than ChatGPT but better than Flux.

It fundamentally reimagined the composition with good enough adherence to instructions.

Still, it introduced unauthorized elements that completely transformed the requested scene—this AI that prioritizes its aesthetic vision over literal instruction following.

It generated a black background, the cat was not correctly placed, there was some color spillage, and elements were not really surreal.

Winner: ChatGPT is by far the undisputed leader in prompt comprehension, accurately rendering complex instructions that caused competing systems to fundamentally break down.

This capability represents a crucial advancement for practical creative workflows where precise visualization of specific concepts is essential. Reve comes second with Flux in a very far third place

Image Editing

ChatGPT’s natural language editing capability represents perhaps its most transformative feature, allowing intuitive modification through conversational instructions while simultaneously providing granular control comparable to specialized tools.

Where traditional image generators often require technical precision or specialized knowledge of plugins, inpainting techniques, etc, ChatGPT’s implementation enables creative experimentation through natural dialogue.

Our tests transforming personal photos into movie posters demonstrated exceptional versatility—a workflow no competing model matched.

For example, we simply fed the model a photo of Decrypt co-founder Josh Quittner and instructed it to generate a Netflix poster with a specific aesthetic, title, and lettering.

It did everything almost flawlessly. Achieving similar results that other models would take a lot of time to undertake, and likely using different tools and plugins.

By the way, this is the feature everyone loved and led to the viral spread of “Ghibli-style” transformations on social media today.

It’s basically a reimagination of a complete scene using simple natural language instructions to generate very complex images.

While all systems eventually show quality degradation through multiple iterations (an expected limitation when regenerating rather than modifying existing pixels), ChatGPT maintained superior image coherence through extended editing sequences compared to both Reve and Gemini.

For example, it still generated coherent, good-quality faces after several iterations, whereas Gemini stopped producing usable results after four or five tries.

Bonus: GPT has a granular “inpainting” feature—allowing you to modify specific areas of an image while seamlessly blending in with the background– for users in need of a more specific editing tool, which Gemini and Reve lack.

Winner: ChatGPT is by far the best model for image editing because it offers natural language understanding and localized inpainting. Reve follows in second place, with Gemini in the third spot due to its quality degradation after several iterations

Content moderation

Despite implementing comprehensive safety measures, our testing identified some vulnerabilities in ChatGPT’s image generation guardrails.

With minimal experimentation, we were able to generate potentially problematic content.

For example, while the system initially refused to generate an image involving a child and substances, it proceeded when prompts were reworded using euphemistic language while maintaining fundamentally identical content.

It would not generate a child inhaling cocaine with a rolled dollar bill, but a child with white powder and a rolled green paper the size of a dollar bill is totally fine.

Try as we might, we were unable to generate overly sexualized photos, violence, and other questionable content simply by convincing the model of our good intentions.

Conclusion

GPT-4o’s image capabilities establish a new benchmark in AI-assisted visual creation—one that combines exceptional technical performance with unprecedented accessibility.

For most users, this implementation now represents the optimal balance of quality, versatility, and value for $20 a month.

Other specialized tools only let users handle text and code, or just images—but you can’t find an all-in-one offer with the same levels of quality making OpenAI’s service not only easy to use but a great value proposition.

Edited by Sebastian Sinclair and Josh Quittner