After trying GPT-4o to draw pictures, I felt for the first time that human designers are a bit dangerous

As the organizer of the GetAI Learning Circle, I demonstrated how to use AI to generate images in the live broadcast on the first day of the circle.

At that time, I also used Midjourney to make a batch of mobile phone case pictures for everyone.

But I have to say that the threshold for AI drawing was still quite high at that time. I wrote down a lot of excellent command words in my little notebook. I love learning and I insisted on going to the official website every day to find a picture made by others and study the secrets in their command words.

But this kind of "memorizing the question bank" effort was actually wiped out in an instant after the leap in AI. Soon, everyone was back to the same starting line.

I have this feeling because on March 26, OpenAI's GPT-4o basic model underwent an update.

To be honest, I didn’t pay much attention to it at first. After all, among OpenAI’s products, whether it’s the o1 that focuses on reasoning capabilities or the o1Pro that costs $200 a month, the attention they receive far exceeds that of the 4o.

But what impressed me most about 4o this time was its ability to produce raw images.

I tried it out for a while and would like to share with you what impressed me most.

The first feeling is that the quality of the picture is really realistic.

I also specifically looked up the pictures I generated using AI for the first time in February 2023.

At that time, I asked it to draw a young man and woman from the 1980s sitting in a square. The final result was so terrible that I could only praise it.

Two years passed in a blink of an eye, and I used the same command words to let GPT-4o redraw the same picture.

When you compare the two pictures together, you will have a sudden realization - it turns out that AI has quietly iterated so much in its own world.

AI Works in 2023

AI Works in 2025

My second feeling is that Chinese can now be generated directly on the picture.

I used to recommend that you use AI to draw a background image before making a poster. But if you want to add a title or slogan to the image, you still have to go to Meitu Xiuxiu or Photoshop to add the text manually.

This time, 4o can generate Chinese characters directly on the picture. Although there are occasional garbled characters, most of the time, the text can be accurately displayed.

When I saw it could write Chinese, my first reaction was: "It would be great if it could change to a better-looking font every time I say something."

You see, human beings are so unsatisfied.

The third feeling is that now you can directly change a photo into the anime style of Hayao Miyazaki, or change it to another style.

If you have been browsing your friends circle these days, you have probably seen a lot of similar anime pictures, which are most likely made using the new 4o model.

I couldn't help it and tried it with a photo of me, Luo Pang and Tuo Buhua. The effect was quite interesting.

The fourth feeling is that it can now change the picture with just one sentence.

This feature really caught my eye. Because GPT-4o has a strong ability to understand, as long as you make your requirements clear, it can understand and start working immediately.

I tried a few very everyday scenes and the results were very good.

For example, I throw a picture in and say "change the color of the clothes to blue", and it changes immediately.

For example, if I throw two pictures in and say "merge", it will directly merge them into a new picture for me.

Another time, I took a picture of a desk and asked it to cut out two figurines on the desk. It turned around and made me a picture with a transparent background, with only the two figurines.

This means that in the future, when it comes to things like making advertising posters, as long as you explain it clearly and the AI ​​can understand it, it can do it directly with incredibly high efficiency.

Of course, AI doesn’t always succeed.

For example, one time, I asked it to change a photo of three people into Argentina uniforms, and it told me directly that this task was a bit difficult.

I asked it why it couldn't do it, and it said that in this picture, three people were all wearing white shirts and the background was also gray and white, so it was a bit difficult to change it accurately.

Then it gave me two plans very seriously: one was to take out the three people separately, change their clothes separately, and then put them back together; the other was to change one person's clothes first, send it to me to see the effect, and if it was OK, it would continue to change the other two.

To be honest, when I saw its response, I was really worried for many designers who were not particularly outstanding. Let’s not talk about the technical level and work speed, just its communication ability alone is better than many human designers.

The improvement of 4o in raw images is not entirely due to the AI ​​itself becoming stronger. More importantly, it is assisted by reinforcement learning with human feedback.

OpenAI has hired a team of hundreds of trainers to stare at the AI-generated images, annotating typos, distorted details, or deformed hands and feet one by one. Then, through reinforcement learning, it continuously trains the AI ​​to learn how to understand human language more accurately and do things accordingly.

After talking about the four major feelings it brought me, let me talk about the changes that may occur next.

The first change is that the SD image workflow, which has been extremely popular in the past two years, may gradually lose its popularity.

In the past, institutions like photo studios and schools needed to use it, but the cost was high, it was difficult to get started, and the threshold was not low. But now that the 4o version has come out, some AI intelligent agent workflows that pursue simple effects suddenly seem less necessary.

Judging from this detail, the AI ​​workflow we are currently studying is actually probably just a transitional solution. There will definitely be small tools to solve specific problems, but they don’t necessarily have to use the complex method of “workflow”.

Because now, no matter whether it is the buttons or the intelligent bodies made in other places, they are mostly used by individuals to play with themselves or demonstrate their abilities, and it is difficult to truly embed them into the workflow of the enterprise.

The second change is that it is particularly friendly to many professionals with ideas.

For example, if you want to make a poster or a logo in the workplace, or if you are a school teacher want to draw a courseware illustration, or make a schematic diagram of a physics experiment, these can now be directly handed over to AI.

Moreover, as the quality of AI images becomes higher and higher, we can also continue to throw the drawn pictures into AI video tools such as Keling and Hailuo, and let them automatically generate short videos.

I tried it myself. I asked GPT-4o to draw a picture of the entrance of the Peach Blossom Spring cave, and then threw the picture into KeLing, which turned into a short video in a few minutes.

The third change is that the cycle of many projects will be greatly compressed in the early stages.

Whether it is product development in an enterprise or a concept proposal in an advertising company, everyone can now let AI draw their ideas first, and then the team can look at the pictures and discuss them together.

In this way, in addition to exchanging ideas, more new creativity can be stimulated and efficiency is greatly improved.

The fourth change is that the platform is really friendly to those self-media students who write public accounts.

In the past, it took me a long time just to find a matching picture. If it was just a waste of time, it would be fine, but the most frightening thing was that I used a picture that shouldn’t be used, and was targeted by some unscrupulous organizations and blackmailed for copyright.

But now it’s different. As long as you have a picture in your mind, just describe your idea to GPT-4o, and it will help you draw it. If you don’t even have a sense of picture, just throw your written text in and let it automatically match pictures according to the content.

I have also tried this scene myself, and the effect picture is at the end of the article. You can experience it for yourself.

I also checked my Midjourney bill and found that my first payment was $10 on March 17, 2023, and I have been paying for two years. But to be honest, this 4o update is the first time I have seriously thought about whether to cancel Midjourney.

Then, the classic line popped up in my mind: " When you used to watch the moon with me, you called me Little Sweetie; now that the new one is better than the old one, you call me Mrs. Niu. "