Google has recently introduced an updated version of its Imagen 3 AI image generator within the Gemini platform.
With the latest release of Imagen 3, Google is committed to delivering more accurate and visually appealing images based on user text prompts. Notably, the AI can now create images of people again, though significant restrictions are still in place. The model is designed to avoid producing photorealistic images of identifiable individuals, such as celebrities and politicians.
Additionally, it will automatically block content that is “excessively gory, violent, or sexual,” ensuring that the generated images adhere to ethical standards.
Imagen 3, like other generative AI models such as OpenAI’s Dall-E 3 and Stable Diffusion, operates by transforming text prompts into visual outputs. However, shortly after its initial launch, users observed that Google’s efforts to filter out stereotypes may have been overly aggressive. For instance, prompts that typically depicted lighter-skinned men occasionally resulted in images of darker-skinned individuals, sometimes even women, leading to a significant backlash.
One example involved users asking the model to depict a U.S. Senator from the 1800s. The AI consistently produced images of older white men, rather than a diverse array of individuals. This misstep invited criticism online, with figures like Elon Musk accusing Google of being “anti-civilizational.” In response to the growing discontent, Google temporarily halted the human image generation feature in February to address these issues. During this pause, attempts to generate images of people using Gemini resulted in error messages.
The updated Imagen 3 model will initially be accessible to paying Gemini users, including those with Gemini Advanced, Business, or Enterprise accounts. However, as of this update, the feature is not yet fully live. Users can test the new model in Google’s AI Labs, where early impressions suggest that the AI is less reluctant to generate images of white men. Despite these changes, Google maintains that it continues to adhere to its core design principles, and all generated images are marked with its SynthID watermarking system to ensure they can be identified as AI-created.