During a livestream on Tuesday, OpenAI CEO Sam Altman reported a significant new update to ChatGPT’s image-generation capabilities, marking the first major enhancement in over a year.
ChatGPT is now equipped to use OpenAI’s GPT-4o model to generate and modify images and photos directly. Although the GPT-4o model has been the foundation of the AI-powered chatbot platform for some time, it has previously been limited to generating and editing text alone, without support for images.
Altman indicated that the GPT-4o’s native image generation feature is currently available in ChatGPT and Sora, OpenAI’s AI video-generation product, for subscribers of the company’s $200-a-month Pro plan. OpenAI plans to extend this feature to Plus and free users of ChatGPT, as well as developers utilizing the company’s API service, in the near future.
The image output capability of GPT-4o “thinks” for a longer period than the image-generation model it replaces, DALL-E 3, to produce what OpenAI describes as more accurate and detailed images. GPT-4o is also capable of editing existing images, including those featuring people, by transforming them or altering details like foreground and background objects through “inpainting.”
To enable this new image functionality, OpenAI informed the Wall Street Journal that GPT-4o was trained using “publicly available data,” alongside proprietary data obtained through partnerships with companies like Shutterstock.
Many in the generative AI industry regard training data as a competitive asset and are reticent to disclose information about it due to potential intellectual property-related legal challenges.
In a statement to the Journal, Brad Lightcap, OpenAI’s chief operating officer, assured, “We’re respecting the artists’ rights in terms of how we produce outputs, and we have policies in place to prevent generating images that directly mimic the work of living artists.”
OpenAI provides an opt-out form for creators who wish to remove their works from its training datasets. The company also honors requests to prohibit its web-scraping bots from gathering training data, including images, from websites.
The upgraded image-generation feature of ChatGPT arrives following Google’s experimental native image output for its Gemini 2.0 Flash model. This powerful capability quickly gained attention on social media, albeit not entirely positively. The image component of Gemini 2.0 Flash lacked sufficient safeguards, enabling users to remove watermarks and create images of copyrighted characters.
This article was updated at 12 PM PT to include OpenAI’s statement to the Wall Street Journal regarding GPT-4o’s training data.
