Google is restoring the ability to generate AI images of people in Gemini after temporarily stopping the feature due to inaccurate images.

Google is allowing Gemini to create AI images with humans again. The company confirmed this in a blog. The function is being reopened on a small scale: from today, only companies and paying English-speaking users can use it.

The image generator in Gemini caused a big stir in February. The Imagen model, OpenAI’s counterpart to DALL-E, generated images of people that were not historically accurate. When asked to create “a German soldier from 1943,” a black man and an Asian woman were presented, a generated image of the American founding fathers featured black men, and an image of the Google founders was given an Asian flair, to name just a few. to name just a few examples of the mistakes Gemini made.

Google had deliberately trained the Imagen model with diversity in mind, to avoid images that reflected “a range of people.” However, the model also extended this diversity to images where it was completely inappropriate in a particular cultural or historical context. Google has temporarily suspended the ability to create images of people and has publicly apologized.

Learn from mistakes

To prevent it from going wrong the second time, Google says it has made technical improvements. Limitations are also being introduced. For example, Gemini will not respond to requests to imitate real people or to take explicitly violent or sexual images of people. Google warns that despite these corrections, the tool will not be perfect.

Google is updating the model behind its image generation capabilities. Imagen 2 makes way for Imagen 3. The new model delivers higher image quality and can also learn multiple art and photography styles.

Google is also launching Gems today, modified versions of the Gemini chatbot. The feature is available to paying users.