Gemini AI is getting a new image generation model called Imagen 3, several months afterGoogledeactivated its predecessor due to its controversial output. The search giant announced in itsThe Keywordyesterday, saying, “Imagen 3 brings advanced image generation capabilities that come with built-in safeguards and adhere to our product design principles.”
Google’s Imagen 2 image generation feature in Gemini gathered headlines in February when it produced highly inaccurate portrayals of historical subjects. The company wanted its AI image generation tool to avoid the pitfalls of bias and depictions of real people. However, in an effort to be more inclusive, the tool that it created went too far into the other end of the spectrum, where it had some biasagainstlight-skinned people.
This problem became viral with the post of X user End Wokeness, where the results for “a portrait of America’s Founding Fathers” showed a Native American man, an Asian man, and two dark-skinned men. It even went to the point that the image generator made grossly inaccurate portrayals, where people of color were portrayed as Nazi soldiers.
America’s Founding Fathers, Vikings, and the Pope according to Google AI: pic.twitter.com/lw4aIKLwkpFebruary 21, 2024
New AI technologies must contend with these problems, especially as these programs have no comprehension of human history and ethics. Because of this, AI tools, likeMicrosoft’s Copilot Designer, could produce inadvertently reprehensible images. Google attempted to fix this by ensuring its image generation tool avoids these cultural stereotypes and biases. However, one of its unintended consequences is that the Imagen 2 tool became too inclusive to the point that it became historically inaccurate and, sometimes, grossly offensive. Because of this controversy, Google had to pull Imagen 2 three weeks after its launch to fix it.
The company said it would relaunch the image generation tool in a few weeks, and it finally did so nearly 20 weeks after the pull-out. According to The Keyword blog post, Imagen 3 does “not support the generation of photorealistic, identifiable individuals, depictions of minors, or excessively gory, violent, or sexual scenes.” Nevertheless, it acknowledges that the generative AI tool might make mistakes, but Google says it will continue improving it through user feedback.
Imagen 3 will start rolling out to Gemini users in the coming days, with the Gemini Advanced, Business, andEnterpriseusers in English among the first to get access. With the tool having received technical improvements and better data sets, as well as having gone through intense testing and given clear product principles, we hope to get better, more accurate, and less offensive results.
Get Tom’s Hardware’s best news and in-depth reviews, straight to your inbox.
Jowi Morales is a tech enthusiast with years of experience working in the industry. He’s been writing with several tech publications since 2021, where he’s been interested in tech hardware and consumer electronics.
