OpenAI Breakthrough: GPT-4o Enhances Realistic Image Creation in ChatGPT
March 26, 2025
OpenAI made a big improvement to its AI system by updating its image creation model, GPT-4o. The new model can generate images that look more realistic and also produce clear, easy-to-read text. This update comes after a year-long effort where human workers helped train the system.
What’s New in GPT-4o?
New Default Model
GPT-4o has replaced the older DALL-E 3 as the main image generator for ChatGPT. Now, anyone using ChatGPT—whether on the free version, Plus, Team, or Pro—can use it.
Multimodal Abilities
The previous version of GPT-4o was a multimodal model. This means it could work with text, videos, audio, and images. The refined version makes it even easier to create realistic images and text.
Practical Uses
With the new updates, users can now easily create things like company logos, slide decks, and more detailed images combined with text.
How Did They Improve It?
OpenAI used a method called “reinforcement learning from human feedback” (RLHF). In simple words, over 100 human workers carefully checked and labeled the data used to train GPT-4o. They pointed out mistakes like typos or errors in the images (like extra hands or faces) so that the AI could learn to follow instructions better and create more accurate images.
Gabriel Goh, the lead researcher on the project, explained that the base AI is already smart, but the human feedback made it even smarter and better at producing high-quality images and text.
Impact and Real-Life Examples
Business Use
The improvements in GPT-4o mean that ChatGPT is now more useful for businesses. For example, the new model can create images with transparent backgrounds. This is especially helpful for designing logos and icons. Jackie Shannon, an OpenAI product lead, mentioned that businesses can even ask ChatGPT to generate images that follow a specific brand style.
Customer Example
GoDaddy, a well-known web-hosting company, uses GPT-4o to help create stock images and logos. Travis Muhlestein from GoDaddy said that the new technology is a great tool for embracing AI-driven content creation.
Some Limitations and Controversies
Even though GPT-4o is a major improvement, it isn’t perfect. For example, in one test, a user uploaded a picture of their living room that had two windows. When the AI recreated the living room with new furniture, it only showed one window instead of two.
There is also controversy around AI image generators. Some artists worry that these tools might copy their work without permission and harm their ability to earn money from their art.
Data and Copyright Policies
OpenAI trained GPT-4o using data that is publicly available, along with special data from partnerships with companies like Shutterstock. Brad Lightcap, OpenAI’s chief operating officer, assured that the company respects artists’ rights and has rules in place to avoid creating images that copy the work of any living artist.
This breakthrough shows how OpenAI is working to make AI tools even more useful, whether for everyday users or businesses looking to create high-quality content.