Actual test of OpenAI's latest image generation model. Netizens: It's over.

The AI flavor of GPT Image 1.5 is still a bit strong.

Is the alternative to OpenAI's Nano Banana Pro here?

According to a report by Zhidongxi on December 17th, today, OpenAI launched its new - generation image model GPT Image 1.5. This generation of the model has stronger instruction - following capabilities, more precise image editing, can better preserve details, and its generation speed is four times that of the previous generation model.

OpenAI officially showcased the model's capabilities in a promotional video. It can be seen that GPT Image 1.5 accurately integrated the figures in the picture into different backgrounds such as space and rainforests, and maintained the consistency of the figures among different styles like hand - drawn and felt.

Meanwhile, OpenAI also launched an independent image generation section in ChatGPT, providing various templates and styles to make creation more convenient.

These updates have unlocked many new ways of playing. Sam Altman, the founder and CEO of OpenAI, showed off his "firefighter photo calendar" created with GPT Image 1.5.

However, some netizens found that the calendar in the picture was inaccurate, and many netizens advised Altman to delete the rather eye - catching photo. The official account of ChatGPT couldn't help but use GPT Image 1.5 to put a T - shirt on Altman.

OpenAI has not officially announced any benchmark tests yet. However, on the authoritative large - model evaluation website Artificial Analysis, GPT Image 1.5 topped the two lists of text - to - image generation and image editing, surpassing Google's Nano Banana Pro in both.

In the LMArena large - model arena, GPT Image 1.5 also dominated the two lists of text - to - image generation and image editing.

GPT Image 1.5 is priced by tokens, and the price depends on the resolution and quality settings. The price of one million - pixel high - quality images is about $133 per thousand (approximately RMB 937), and for low - quality images, it is $9 per thousand (approximately RMB 63). All ChatGPT users can use this model today, and its API has also been launched simultaneously.

What is the actual technical strength of GPT Image 1.5? After its release, many netizens have compared the generation effects of GPT Image 1.5 and Nano Banana Pro, and Zhidongxi has also experienced the capabilities of the two models.

Our feelings are similar to those of many netizens: although GPT Image 1.5 is a good image - generation model, there still seems to be an obvious and perceptible gap between it and Nano Banana Pro in terms of realism and detail accuracy.

01. The generation effect has an obvious "greasy feeling", and netizens claim that OpenAI "is completely finished"

First, let's take a look at the text - to - image generation ability of GPT Image 1.5. Our first prompt examined the model's performance in complex scenarios and multi - subject relationships:

A hyper - realistic style picture: On the rainy Tokyo street at night, the neon lights are reflected on the wet road surface. In the foreground, there is a young woman wearing a transparent raincoat, holding a glowing holographic umbrella; in the middle - ground, there is a slowly moving taxi, and the side face of the driver can be seen through the car window; in the background, there is a city skyline with high - rise buildings and a blurred crowd. Cinematic composition, shallow depth of field, 4K details.

In terms of generation speed, Nano Banana Pro is better. It takes about 15 seconds to complete the output, including its thinking process. It can be seen that Nano Banana Pro accurately restored the details in the complex prompt, but did not understand the detail of the "glowing holographic umbrella". In the picture, the details of the taxi and the street were accurately restored.

Subsequently, GPT Image 1.5 also gave the generation result. At first glance, we could feel the obvious "AI flavor". The painting style of GPT Image 1.5 is very "greasy", with a very high saturation. Regarding our clear requirement of "the side face of the driver can be seen through the car window", GPT Image 1.5 blurred it. The integration of the figures and the background is not natural, giving a feeling of being on two different layers.

If you look closely, you can also find that the character in the picture has only four fingers on the right hand. This basic human - body error is really inappropriate for an image - generation model in 2025.

The next prompt mainly examined the model's performance in style transfer and semantic constraints:

Use the brushstrokes and color style of Van Gogh's "Starry Night" to depict the interior hall of a futuristic space station: Outside the huge curved glass window, there are rotating nebulae and planets. Inside the room, there are three astronauts floating and operating the holographic interface in a low - gravity environment. Keep the strong swirling brushstrokes, but with a clear structure and distinguishable objects.

The generation of GPT Image 1.5 was still a bit slow. But let's first look at its effect: although the content of the picture is basically accurate, in terms of the most crucial strong swirling brushstrokes and color style, it is just barely satisfactory, and the difference from Van Gogh's original "Starry Night" is very obvious.

The generation result of Nano Banana Pro is as follows. It can be seen that on the premise of ensuring accurate details, the model accurately restored the painting style of Van Gogh's "Starry Night", and the color style is closer to the original work.

This prompt mainly examined the consistency of details. The perspective of the picture is also an unconventional one, which can reflect the model's ability in edge scenarios:

The view from a cat's first - person perspective: In the early - morning kitchen, the sunlight shines in obliquely through the window. There is a cup of steaming coffee and a bitten piece of bread on the table. The cat's front paws and the edges of its whiskers are faintly visible at the bottom of the picture. Wide - angle lens, warm color tone, lifestyle photography style, high - detail real texture.

GPT Image 1.5 seriously malfunctioned in this edge scenario. First of all, the cat only has whiskers on half of its face, and details such as the nose are all missing, making it hard to tell whether it is a cat's face or a small fur ball at first glance. In addition, the blurring of the background actually makes the realism of the image worse.

The generation result of Nano Banana Pro is as follows. From the picture, it can be immediately judged that this is the first - person perspective of a cat we requested, and the restoration of light and shadow and the presentation of details also meet our requirements.

Many netizens have also shared the results of comparative tests. For the same portrait, in the picture generated by GPT Image 1.5 on the left, the character's head is too large, and the daily - life feeling of the light and shadow effect is a bit worse. Although the face lighting in the generation result of Nano Banana Pro is a bit insufficient and the window is a bit over - exposed, it is precisely these flaws that make the image more realistic.

The netizen who shared this generation result said: OpenAI is completely finished.

However, some netizens added that if requirements such as "unprocessed iPhone photos" and "low - saturation color profile" are added to the prompt sent to GPT Image 1.5, its effect can be made more realistic.

The AI blogger Heisenberg shared the recently popular giant special effect. He believes that, in comparison, the result of Nano Banana Pro is much more natural. In terms of details, there are many bugs in GPT Image 1.5, such as two cars on the left driving directly face - to - face, and the white lines on the road are intermittent, and Altman's hand also seems too large.

We also tested the ability of GPT Image 1.5 to generate Chinese. The model maintained relatively high accuracy in the first few characters, but many errors appeared later.

02. Supports multi - element fusion and editing, and provides pre - made stylized templates

Currently, GPT Image 1.5 provides about 5 free trial opportunities per day. We were unable to conduct actual tests on the image - editing task. However, OpenAI shared many cases on its official blog.

GPT Image 1.5 supports multi - element fusion. For example, in the following picture