OpenAI's all-new GPT Image makes a surprise debut, thoroughly outclasses Image 2, and the plastic-like artificial texture has finally disappeared.
OpenAI is about to drop another bombshell!
Today, a brand new image model with the code name mona-lisa-1 officially debuts on Arena.
Netizens who tested it firsthand are amazed that the image generation effect of mona-lisa-1 is absolutely phenomenal, visibly surpassing GPT Image 2 to the naked eye.
The generated images look more realistic, with the obvious plastic-like texture greatly reduced.
The new king of GPT Image is coming
Four months ago, GPT Image 2 made a stunning debut and instantly took the AI community by storm.
Back then, Image 2 completely outperformed Nano Banana Pro by a large margin and has remained the top of the AI image generation field ever since.
But now, the landscape is changing.
An anonymous model with the code name "mona-lisa-1" has quietly entered the LM Arena blind test ranking list.
Netizen @AiBattle_ did in-depth investigation, threw the generated image of this model into the OpenAI verification tool for checking —
Surprisingly, the SynthID watermark was detected immediately! The case is closed, this mysterious new breakthrough is also released by OpenAI.
It seems that this time, OpenAI is determined to push itself to the extreme and outperform its own existing products.
The plastic-like texture is gone
Developer Riccardo Wolf fed the same prompt to these two models respectively.
It can be seen from the following outputs that mona-lisa-1 has better realism and less plastic texture.
From left to right: Image 2 (medium); mona-lisa-1; Image 2 (high)
Take the generated real-person photo of Ultraman as an example, the improvement of texture is particularly obvious.
In the view of netizen Tim Jayas, the output of mona-lisa-1 is far richer in details.
In addition, when generating this kind of artistic painting with profound artistic conception, mona-lisa-1 also shows extremely high aesthetic level.
No matter the color rendering, the layers of light and shadow, or the creation of the overall atmosphere, it completely gets rid of the rigid "algorithmic feel" that used to be common in AI painting.
Infographics, more complex
GPT Image has long been good at generating complex infographics.
In actual tests, someone used mona-lisa-1 to directly generate a web UI with extremely realistic visual effect.
For the "human organ exploded view" that extremely tests detail processing capability, mona-lisa-1 also handles it with ease.
Whether it is the "GPT-5 one-page overview" with high information density, or the imaginative "ChatGPT exclusive self-portrait", it can generate them accurately.
It can also imitate Einstein's handwriting and output the finished work directly.
Some people say that the new ChatGPT image model has once again ended the competition.
A simple prompt can generate a two-dimensional wallpaper.
There are tons of amazing demos, just look at the pictures below:
The "core architecture" remains unchanged
Training was completed last year
Careful developers found that the training of mona-lisa-1 was very likely to be finished last year.
@ivanainai conducted 5 consecutive tests related to dates, and all responses of the model were stuck at the year 2025 without exception.
Later, she tried to make the model process content involving the latest news, all attempts ended in failure. Of course, to a certain extent, this is also because this anonymous version does not have internet access enabled.
It is worth noting that if we make a horizontal comparison: the knowledge cutoff date of GPT Image 2 is exactly December 2025.
That means, judging from the limited information revealed so far —
mona-lisa-1 is more like a