On Tuesday, OpenAI unveiled the enhanced image generation capabilities integrated into its GPT-4o artificial intelligence (AI) model. The San Francisco-based company introduced the 4o Image Generation model, emphasizing utility over aesthetics. This new feature includes accurate text rendering, rigorous adherence to prompts, consistent character depiction, and the ability to edit images using text commands. OpenAI has also implemented measures to address the risks associated with deepfakes and harmful content creation.
ChatGPT Introduces Upgraded Image Generation Features
Prior to this update, ChatGPT had the ability to generate images through one of its DALL-E models, but the experience was limited in character consistency and text generation quality. In a blog post, OpenAI announced plans to elevate image generation to a core function of its language models.
Image generated using GPT-4o
Photo Credit: OpenAI
By enhancing its large language models (LLMs), the company asserts these models will now naturally generate and modify images. The extensive parameter size and post-training adjustments allow for a better understanding of context, enabling the models to accurately fulfill user requests. Additionally, as language models, they are adept at producing and rendering text with precision.
The updated image generator has been trained on the correlation between online images and text. OpenAI claims this new system offers improved character consistency, allowing users to generate multiple images of the same character with minimal effort in re-specification.
Images with text generated using GPT 4o
Photo Credit: OpenAI/Derya Unatmaz and Les Morgan
Moreover, the model can generate images containing a significant amount of accurate text, such as signboards, menus, and whiteboard notes. Users have the option to upload an image as input, and the chatbot can recreate it in various styles while allowing for edits.
With the latest image generation enhancements, ChatGPT will also support multi-turn generation. Users can prompt the AI chatbot for modifications or additions to a generated image, and it can refine the output while maintaining the integrity of other elements. OpenAI has indicated that the model can accurately manage between 10 to 20 distinct objects within a single image.
Photorealistic image generated using GPT-4o
Photo Credit: OpenAI
These advanced features are currently accessible to ChatGPT Plus, Team, and Pro subscribers. Despite earlier availability to free-tier users, OpenAI CEO Sam Altman announced in a post on X (formerly Twitter) that the rollout to free accounts will be paused due to overwhelming demand.
In an interesting development, users have taken to social media to share Ghibli-style recreations of their images and popular memes generated using GPT-4o, leading to a surge in Ghibli-related content on the platform. Altman even changed his profile picture on X to a Ghibli-influenced version of himself.
Regarding safety measures, OpenAI is integrating Coalition for Content Provenance and Authenticity (C2PA) information into the metadata of all AI-generated images, enabling easier differentiation from authentic content. The company has also created an internal search tool to verify images produced by its models.
Additionally, measures are in place to block requests for generating images that feature harmful content, such as child sexual abuse material and sexual deepfakes. When users edit images of real individuals, restrictions have been implemented to control the types of imagery that can be generated.
