Here is the rewritten text:

The OpenAI has announced a new 4th generation model for image generation, already integrated into the ChatGPT. This innovation produces more realistic results and stands out for its ability to render and place legible texts in creations.

The use of text in AI-generated images has always been a critical factor: models from different companies had difficulty reproducing all letters and left distortions along the way. “The result is an image generation that is not only beautiful but also useful,” said the creator of ChatGPT in a statement.

Learn more about the new AI:

What the new model can do?

The 4th generation model is more capable of understanding the context of prompts, according to the company, allowing for the creation of images with texts or rich in details. The level of realism is higher, and there are more options to refine a creation.

Beyond texts, it is possible to send more instructions to the model: ChatGPT was able to create images containing 15 different objects during a demonstration.

However, the company warns that the technology has difficulty following 20 or more instructions in the same command. It would not be possible to create a periodic table or a list of NBA teams with precision, for example.

The novelty also allows for creating images with transparent backgrounds and transforming them into stickers that can be used in other creations. It is also possible to change the style of a creation without losing essential elements, including an option to modify an existing image.

The OpenAI informs that all creations with the new model follow the guidelines of the C2PA (Coalition for Content Provenance and Authenticity, or “Coalition for Content Provenance and Authenticity” in free translation), a standard for invisible watermarking to signal that something was generated by AI.

Limitations

The developer also explained that the model has some limitations. See examples:

• Very long images may be cut off at the bottom;
• The model may hallucinate and invent incorrect information, mainly in prompts with little context;
• The model may have difficulty creating images with more than 20 concepts at once;
• The AI has difficulty creating precise graphics;
• The model may show errors when rendering characters that are not from the Latin alphabet;
• Adjustments to small areas may show errors;

The OpenAI also reinforced that it will block requests for images that may violate the company’s content policies, including erotic deepfakes, and applied restrictions for creations involving the appearance of real people. The company did not comment on the use of copyrights or materials used for training.

Availability

The new model began to be released on Tuesday (25) for free users and subscribers of the Plus, Pro, and Team plans of ChatGPT. In a test done, the chatbot still used DALL-E to generate images, and the 4th generation model was not available.

Read also:

• 7 free alternatives to ChatGPT
• How to create a prompt in ChatGPT?
• How to know if a photo was created by Artificial Intelligence

VIDEO: How to use ChatGPT on WhatsApp