DALL-E 3 was released back in September 2023, and it’s been a year and a half since then. For the AI era, that update pace is indeed a bit slow, and during this time, many users have been asking in OpenAI’s official forums when a new image generation model would come out. Finally, just a little while ago, OpenAI launched the new generation 4o Image Generation feature. That’s right—now GPT-4o can generate images too, and it’s not just for paying users; the first wave has also been opened up to free users.
More importantly, this time 4o’s image generation also supports text — both English and Chinese work. I tested it and the results were fantastic.
OpenAI’s all-new 4o Image Generation feature has arrived
Compared to the previous DALL-E 3 model, which was purely for image generation, the image generation feature now integrated into GPT-4o is even more versatile. Not only can it generate images, but it can also add content to existing images, convert them into different styles, turn code into images, and reference multiple images to generate a new image based on a described request.
OpenAI stated that through native multimodal models,4o Image Generation It enables precise, accurate, and realistic image output, meaning it is no longer just generating “roughly similar” images, but can create image content very close to the real world with extremely high fidelity and detail control.
Below are several images provided by OpenAI.
一張偷拍風格的照片，畫面中 Karl Marx 匆忙地穿過美國商城的停車場，他驚慌地回頭張望，試圖避開鏡頭，不讓人拍下來。他緊抓著數個閃亮的購物袋，裡面裝滿奢侈品，大衣在風中飛揚，其中一個袋子因他正邁步走路而晃動著。背景呈現出模糊的汽車與發光的商城入口，強調動態效果。鏡頭的閃光使得部分畫面過曝，整體照片呈現出一種混亂、八卦小報風格的感覺。
生成一張寫實風格的圖像，場景為 2006 年夏季某個週六在多倫多的農夫市集。天氣晴朗、陽光明媚，是六月下旬的一個美好日子，人們正悠閒地逛市集、購物並吃著三明治。畫面焦點應放在一位年輕的亞洲女孩身上，她穿著牛仔吊帶褲，手中拿著一杯草莓香蕉冰沙輕啜著。其餘背景人物與市集細節可以模糊處理，讓主體更突出。整張圖像應模仿2006 年數位相機所拍攝的風格，呈現出那種略帶噪點與曝光特性的畫質，並在照片一角加入類似沖印相片的時間戳記。畫面比例為 3:2。
一張模糊的老式類比底片照片，畫面為夜晚安靜巷弄中停靠著的一輛汽車。整體氣氛靜謐，畫質帶有底片特有的顆粒感與柔焦效果，彷彿是過去某段時光隨手拍下的片刻。照片創作者：Roope Rainisto
When using AI image generation in the past, many people have likely experienced that the AI would initially generate an image they were fairly satisfied with, but when they wanted to make some adjustments, the AI’s results would gradually deviate from the original image style.
To this end, OpenAI mentioned that since the image generation feature has been integrated with GPT-4o, GPT-4o can continuously create content based on the image and text context in the conversation, ensuring consistency and continuity throughout the creative process.
Additionally, this new image generation feature can also precisely generate images with specified text.
Other tools typically handle only about 5 to 8 objects, while OpenAI claims that GPT-4o can simultaneously process up to 10 to 20 different objects, with tighter binding between object features and their relationships, giving users finer control over each element in an image.
I actually tested it by using Ada’s profile picture, asking to blend it into this image and add the specified text. Below is my prompt.
生成一張融入這個大頭貼，背景有未來感，並加入電腦元素，呈現一種專業 AI 科技的氛圍，圖片上要有「電腦王阿達」字
The result came out really impressive—the profile picture is perfectly replicated, and the Chinese text was added flawlessly:
Then I asked to change it to a wide size, adding “4o Image Generation” below Computer King Ada, and it worked perfectly too. So impressive:
Even generating realistic-style portraits doesn’t have an AI feel to them.
Notably, all images generated by GPT‑4o will embed C2PA metadata to indicate that the image was produced by GPT‑4o.
GPT-4o image generation is now officially available to all users, including Plus, Pro, Team, and free users. Simply enter your image generation prompt in the GPT-4o model:
Source: KOCPC Chinese