Initially, Google Bard would not return images. Instead, you would get the familiar response of: I’m a large language model, I’m unable to assist with that. However, it now seems it can return images from Google Search.
Simply ask Bard to return images in its response.
However, it seems that Bard cannot generate fake images. The dataset of images that Bard pulls from seems to be well established and properly copyrighted. Google is probably trying to reduce liability by only allowing Bard to use images that have a proper chain of custody.
Input: Show me some photos of bean plants.
Each image is overlaid with a copyright and even contains a link to the source website.
This is another step in the right direction to making Bard more useful.
Of Course, Bard is still makes mistakes:
Input: What is on the dark side of the moon? Provide images in your response.
In this case, Bard utilized Wikipedia to provide an answer, which seems to have combined information about the ACTUAL ‘dark’ side of the moon (which is the side always facing away from the Earth) with the Pink Floyd album titled “The Dark Side of the Moon“. The image that Bard included in its response was the image on the cover of Pink Floyd’s album.
Let’s refine our prompt and try again.
Input: Show me images of the side of the moon facing away from Earth.
Can Bard Generate Images?
It seems Bard still cannot generate or leverage other online services to generate images of things that do not exist. For example, I asked Bard to generate an image of a cat.
I wanted to see if Bard would pull images from the images.google.com dataset. For this test, I searched images.google.com for ‘a cat crossed with a sailfish’. It returned many images of sailfish and boats.
I expect, when asking Bard to show me an image of a cat crossed with a sailfish, it would simply return a few images from the images.google.com dataset.
It seems bard is getting smarter each day!