ChatGPT's image generation feature gets a big update: Here's what's new

OpenAI has unveiled a significant upgrade to ChatGPT’s image-generation capabilities, marking the first major enhancement in over a year. The new feature, called "Images in ChatGPT," allows users to generate and modify images directly within the chatbot using the company’s GPT-4o model.

Native Image Generation in ChatGPT

GPT-4o, which has long underpinned OpenAI’s AI-powered chatbot, now extends beyond text generation to include images. This feature is now available to subscribers of ChatGPT’s Plus, Pro, Team, and Free tiers, with free users having a limited daily quota. According to OpenAI spokesperson Taya Christianson (via The Verge), these limits mirror those of DALL-E 3, though they may change based on demand. DALL-E 3 itself remains accessible via a custom GPT.

Unlike its predecessor, GPT-4o "thinks" longer before generating images, resulting in improved accuracy and detail. The model follows an autoregressive approach, sequentially generating images from left to right and top to bottom, rather than using the diffusion-based technique employed by models like DALL-E 3. This shift may contribute to its improved text rendering capabilities, an area where traditional AI image generators often struggle.

Enhanced Image Editing and Precision

GPT-4o’s ability to modify existing images represents another major leap forward. Users can now alter images - including those featuring people - by inpainting details such as background and foreground elements. This means that images can be refined in real-time through a conversational interface, making iterative adjustments more intuitive.

Additionally, the model boasts superior "binding" capabilities, ensuring it maintains the correct relationships between attributes and objects in a given prompt. Many AI image generators struggle with accurately depicting complex scenes containing multiple elements, typically failing beyond 5-8 objects. GPT-4o, however, can handle between 15-20 objects while maintaining accuracy.

"This model is a step change above previous models," said OpenAI research lead Gabriel Goh while speaking to The Verge. He highlighted how the system improves object-to-attribute binding and text rendering, making it far more reliable for generating structured images with embedded text, such as signs or infographics.

Training and Ethical Considerations

To power this advanced capability, OpenAI trained GPT-4o using publicly available data as well as proprietary datasets obtained through partnerships with companies such as Shutterstock. However, the company remains cautious about revealing too much about its training process, partly due to intellectual property concerns.

OpenAI has also taken steps to address copyright issues, providing an opt-out form for artists who wish to exclude their work from future training datasets. Additionally, the company respects requests to block its web-scraping bots from collecting data, including images, from specific websites.

Despite these measures, GPT-4o-generated images will not feature visible watermarks indicating AI creation. However, OpenAI has confirmed that all generated images will include C2PA metadata to mark them as AI-generated, and the company has internal tools to track images created by its models.

Competitive Landscape

This update arrives amid increasing competition in the AI image-generation space. Google recently introduced experimental native image output in Gemini 2.0 Flash, but the feature quickly drew criticism for its lack of guardrails, enabling users to remove watermarks and generate potentially infringing content. By contrast, OpenAI claims it has stricter safeguards to prevent direct imitation of living artists’ work and copyrighted material.

With these advancements, OpenAI positions ChatGPT not just as a conversational AI but as a powerful multimodal tool capable of seamlessly integrating text, images, and future media formats. As the technology evolves, the ability to generate visually coherent, contextually accurate images within an interactive chat interface could redefine how users create and interact with AI-generated content.