Two days later, a mysterious image-generating model called "mona-lisa-1" appeared on the Arena platform.
Some have called the model "Generating an Image of Your Creator" OpenAI and Ottman.
This is the first time I've seen this.
Some netizens use OpenAI verification tools to detect images that have been generated to confirm the SynthID watermark of an OpenAI image.
Sources: @AiBattle_
Combining these details with naming styles and output features, many have concluded that it may have come from OpenAI, a follow-up to or testing branch of GPT Image 2.
As early as August 4, some netizens noticed that OpenAI was conducting A/B testing, which focused on the ability of prompt words to follow and text rendering. At the time, it was unclear whether this was a brand new model or a regular update to GPT Image2.
OpenAI has not made any official statements about this, but the system is not unknown. Before the official launch of GPT Image2, it was blind-tested in live Arena under the codes Maskingtape, Gaffertape and Packingtape, and was finally officially released in April this year.
Following the release of GPT Image 2, Arena Elo, with 242 credits, led the way across multiple image generating lists to become the most recognized commercial image model.
Today, Mona-lisa-1 has been exposed at the Arena, with some test subjects reflecting an improvement in its richness of detail and sensitivity to sand compared to GPT Image 2, and a certain improvement in its "plastic sensitivity" skin problems that have plagued it for a long time.
Sources: @TimJayas
However, there are also netizens who have had cold water, and some of the generated images still have clear noise and block shadows, and the overall upgrades are limited.
Some netizens have found that the model's knowledge deadline has not been updated and that it does not have internet search capabilities.
This is the first time we've seen this.
What's better than GPT Image 2?
To see how much progress the mona-lisa-1 has made, we can directly compare the two models using the same pronoun.
One netizen used GPT-Image 2 and mona-lisa-1 respectively to create street characters.
Sources: @TimJayas
Both images use similar subjects and structures, but mona-lisa-1 shows more detail, skin texture, facial patches and hair discomfort are more pronounced, and background emptying is more natural.
On the issue of "plastic sensation", Mona-lisa-1 has also seen some improvement.
The testers asked the model to generate a snapshot of the phone and add a series of restrictions to the prompts. The end result was very close to a hand-held photo, less intentional traces, and the five senses were still very detailed, especially the skin's oily glow, redness, and local over-exposure, which made the image look more real.
I'm not sure if I'm going to be able to do that. Harsh direct phone flash illuminates their faces and clothing while the background remains much darker. Slightly awkward framing, one person partly cropped by the edge of the image, minor motion blur, realistic sensor noise, natural skin texture, realistic pores, flyaway hair and ordinary c The photo should look like a completely unedited image from someone's camera roll. No cinematic lighting. No professional photography. No beauty retouch. No smooth plastic skin. No teal-orange grading.
More netizens mentioned that the mona-lisa-1 seems to be better at actively suppressing "AI flavors", reducing overly refined glare, smooth skin and deliberate design.
But it doesn't solve all the problems either.
Mona-lisa-1 can actively generate more skin texture and camera noise, but "more detail" is not necessarily the same as "a more realistic picture".
Some have complained that the model is too fond of writing in pictures.
In order to roughly estimate the deadline for model training data, five successive date tests were conducted, with the mona-lisa-1 each writing the year as 2025.
When asked if the model was still alive, the model gave a clear answer: Charlie Kirk is still alive and there is no credible evidence that he is dead.
Source: @chetaslua.Charlie Kirk was shot and killed during an event on September 10, 2025.
Combined with other date-testing, it has been speculated that the model's training data may be discontinued after May 2025.
It is widely believed that "Mona Lisa" is not related to OpenAI, and the bad news is that it is only a paper advertisement.