Google is working on a text-to-image diffusion model called Imagen. Recently they unveiled it for the world to see a limited part of it. Even though the images displayed had an unprecedented degree of photorealism Google chose not to release the code and the demo. The reason: Society.
Why Google Didn’t Release Imagen
The reason why Google didn’t release a demo or Imagen’s code might sound funny or to some a scene out of Joker’s movie, but Society is the actual reason Google didn’t release its text-to-image generator. Let’s go over the issues one by one explaining each issue:
1. Color Preference/Bias in Representation
Imagen is a text-to-image ai generator where it will have to generate everything including humans. However, it seems that the ai currently prefers to show a certain color. There can be multiple reasons for this, but what might be affecting it the most is its dataset.
The web-scraped datasets sped up the creation process of Imagen by giving them the most of the data required in large quantities with no bias. However, the bias arises from the population difference and general preference of color.
2. Color Preference/Bias in Professions & Activities
The representation of a certain color in a profession and activities is unfair. However, as mentioned earlier, the web-scraped datasets aren’t based on equality, they are based on what people are uploading and viewing.
Google and other search engines can not restrict someone to upload an image of any person portrayed in any position, activity, or profession. So the web scraped data set also portrays certain colors in certain professions, positions, and activities, which is unfair to many.
For example, if you search for the term Gang members the search results say it all. The ratio of light to dark-skinned represented as gang members was 1 to 4.
3. Misuse of Imagen
Imagen created highly photorealistic images, which could be mistaken for real images. In the wrong hands, it could help spread fake news. Google hasn’t come up with a solution to prevent the misuse of Imagen which they should if they actually want to use Imagen. Otherwise, creating such software which can’t be used due to the risk of misusage.
The bias toward people of a certain culture, looks, and color is embedded within our society which is not Google’s fault and in some places anyone’s. Like the population difference between certain colors isn’t something intentional in normal cases. If a certain population is bigger in size then undoubtedly they will be in the mainstream more. For example, if there are six yellow people and four red people and all of them upload their photos on Facebook then the difference in representation or the spotlight will obviously be on the yellow people.
Possibilities of Imagen
The possibilities and applications of Imagen are vast because of its sheer amazing performance and spectacular results. You can see more images that were generated by Imagen on the official website.
The description on which Imagen based the above picture: A photo of a raccoon wearing an astronaut helmet, looking out of the window at night. It is quite impressive what Imagen can build!
Google could clean its datasets and lessen the bias but it would take eternities. What they could do is assign some colors and a probability ratio to humans which could help a lot. But whatever they do we do certainly hope Google succeeds in creating Imagen perfectly because it has shown extremely impressive results till now.
