Google Claims it Fixed Gemini’s AI Image Generation of People

Google is restoring Gemini’s ability to generate pictures of people as it expands its Imagen 3 image generator to apps and all languages.

The company is rolling out person image generation over the course of the “coming days.” Early access will come to English-language Gemini Advanced, Business, and Enterprise users. More users and languages are coming in the near future.

Google said it had addressed the problems by making “technical improvements,” using better evaluation data sets, establishing clear “product principles,” and conducting red teaming exercises that attempted to break the AI model.

Imagen 3 under Gemini won’t produce realistic, identifiable images of people, the company said. It also won’t allow pictures of minors or “excessively gory, violent, or sexual” content. Google cautioned that images wouldn’t be flawless, but that it would also listen to feedback.

The new generative AI system is billed as higher quality overall than its Imagen 2 predecessor. You can create images in specific art styles, and the resulting pictures are closer to what you wrote in your prompt. It supposedly runs “favorably” versus other image generators, according to Product Manager Dave Citron.

Google pulled Gemini’s ability to generate people in February after users produced historically inaccurate images, such as people of color as German WWII soldiers. CEO Sundar Pichai reportedly told staff the results were “unacceptable,” and Senior VP Prabhakar Raghavan said that the model didn’t account for situations where diversity isn’t realistic.

It’s too soon to say if Google has completely resolved the issues with Imagen. Many of the issues found in generative AI models have only crept up as larger audiences got to try them and find weaknesses.

The approach nonetheless contrasts sharply with xAI’s Grok. The Elon Musk-backed generator has relatively few guardrails and lets people create deepfakes that could be used to mislead others. Google is determined to have a safe approach to creating people, in part because major features like the Pixel 9’s Pixel Studio revolve around Gemini.