How To Generate Images With OpenAI’s ChatGPT-4o AI Model?

OpenAI has introduced a major upgrade to ChatGPT, significantly enhancing its image-generation capabilities. The newly launched feature, “Images in ChatGPT,” marks the first substantial improvement in over a year and allows users to create and modify images directly within the chatbot using OpenAI’s latest AI model, GPT-4o.

Native Image Generation in ChatGPT

GPT-4o, the foundation of OpenAI’s AI-powered chatbot, now extends beyond text-based responses to include advanced image generation. This feature is available across different ChatGPT subscription tiers, including Plus, Pro, Team, and Free, though free-tier users will have a limited daily quota.

OpenAI spokesperson Taya Christianson, speaking to The Verge, noted that the image-generation limits for free users align with those of DALL-E 3, though these limits may change depending on demand. While DALL-E 3 remains accessible as a separate model, GPT-4o integrates image generation seamlessly within the chatbot’s interface.

Unlike its predecessors, GPT-4o employs a different approach to image creation. Instead of using a diffusion-based technique like DALL-E 3, it generates images autoregressively—building them sequentially from left to right and top to bottom. This method results in greater accuracy and detail, especially when rendering text within images, an area where traditional AI image generators have often struggled.

Enhanced Image Editing and Precision in ChatGPT-4o

A key advancement in GPT-4o is its ability to modify existing images, including those featuring people. Users can refine images in real time through a conversational interface, adjusting elements such as backgrounds and foregrounds with precision. This “inpainting” capability allows for more natural and iterative edits, making the tool more intuitive for creative work.

Additionally, GPT-4o significantly improves its ability to maintain the correct relationships between multiple objects and attributes in an image. Traditional AI image generators often fail when depicting complex scenes with more than 5-8 objects. However, GPT-4o can handle between 15-20 objects while maintaining high accuracy.

“This model is a step change above previous models,” OpenAI research lead Gabriel Goh told The Verge. He highlighted that GPT-4o has superior object-to-attribute binding and enhanced text rendering capabilities, making it more reliable for generating structured images, including those with embedded text such as signs and infographics.

Training and Ethical Considerations

To achieve this level of sophistication, OpenAI trained GPT-4o on publicly available data, as well as proprietary datasets obtained through partnerships with companies such as Shutterstock. However, the company remains discreet about the specifics of its training process due to concerns around intellectual property and competitive advantage.

Recognizing potential copyright issues, OpenAI has implemented measures to allow artists to opt out of having their work included in future training datasets. The company also honors requests to block its web-scraping bots from collecting data from specific websites.

While GPT-4o-generated images will not carry visible watermarks, OpenAI has confirmed that all images will include metadata from the Coalition for Content Provenance and Authenticity (C2PA) to indicate their AI-generated origin. Additionally, OpenAI has internal tools to track images created by its models, ensuring transparency and accountability.

Also Read: Signal Chat Leak: Mike Waltz Takes Responsibility For Exposing Trump Officials’ Yemen Strike Plans