Google explains what went wrong with Gemini’s image generation feature

Google Gemini made headlines earlier this month for removing the Bard branding and simultaneously replacing the Assistant on Android devices. But despite a promising start to the month, things seem to have fallen off a cliff for Google, especially when it comes to image generation. After hearing more comments from users over the past week about Gemini’s tendency to produce inaccurate images of people, Google decided this week to pause the feature until further notice. Did. The company is now explaining what went wrong and what it plans to do next.

related

5 Google Gemini prompts that make the most of generative AI

Stop boring emails and tedious tasks.These Gemini prompts will change the way you work

Prabhakar Raghavan, Google’s senior vice president of knowledge and information, wrote a detailed blog post about the issue at hand, while also acknowledging that Gemini’s image generation capabilities were “missing the mark.” Commenting on the reason behind the issue, Raghavan said,

First, Gemini’s adjustment to show different people clearly failed to explain cases where it shouldn’t show a range. And second, over time, the model becomes much more cautious than we intended, refusing to answer certain prompts outright, and erroneously labeling highly unusual prompts as sensitive. I interpreted it.

The combination of these two key issues has resulted in Gemini overcompensating in some areas while being overly conservative in other demands, resulting in an “embarrassing and false” image. Raghavan says. The company details the shortcomings of his Gemini, specifically Imagen 2, which the chatbot uses for image generation, but there is no clear mention of when this feature will return to the chatbot.

Google is currently working on “significantly” improving user-generated image generation before moving it back to Gemini, according to a blog post. So it’s safe to say that Google Gemini won’t be able to generate images of people for the foreseeable future, although how long that will take is anyone’s guess. To the company’s credit, we have always maintained that LLMs (Large-Scale Language Models) can make some mistakes, and we are referring to what is widely known in the industry. hallucination. Raghavan concluded by saying that Google will do its best to provide AI to users “safely and responsibly.”

Gemini can still generate images of animals, but not when humans are involved

However, not all types of image generation capabilities remain from Gemini, and users can still generate photos of animals and inanimate objects. However, you cannot retrieve images of people using a general query like “.Generates an image of a person walking a dog.“However, if the request does not include people, as in the example above, Gemini will happily generate a decent image.