DALL-E, a program developed by commercial AI lab OpenAI, has been the field's leader to date (and updated just back in April). However, Google recently announced Images, its own take on the genre, and it has just surpassed DALL-E in terms of quality.
The best way to grasp these models' incredible capabilities is to simply look at some photographs they can produce. Above are some Images-generated samples, with more below (you can see more examples on Google's dedicated landing page).
In each case, the prompt was input into the program at the bottom of the image, and the image above was the output—just a reminder: that's all there is to it. You put in what you'd want to see, and the application creates it for you. Isn't that incredible?
While the coherence and precision of these images are unquestionably astounding, they should also be treated with a grain of salt. When Google Brain or other research teams announce a new AI model, they often cherry-pick the best results. So, while these images appear to be perfectly polished, they may not represent the Image system's average output.
Images generated by text-to-image models frequently appear unfinished, smeared, or hazy, a problem we've encountered with OpenAI's DALL-E software.
Check out this interesting Twitter thread about DALL-E difficulties for more information on text-to-image system issues. It demonstrates the system's tendency to misread instructions and struggle with both text and faces, among other things.
However, based on a new test it built for this project called Draw Bench, Google asserts that image routinely delivers better images than DALL-E 2.
Draw Bench is a simple metric: it's a set of 200 text prompts that Google's team inputs into images and other text-to-image generators, with the output from each program, then rated by human raters. As illustrated in the graphs below, Google discovered that humans prefer image output to that of competitors.
However, we won't be able to judge this for ourselves because Google isn't making the image model available to the public. There's also a valid explanation for this. While text-to-image models offer a lot of creative possibilities, they also have a lot of negative uses. Consider a machine that can generate almost any image you want, and then use it for things like fake news, hoaxes, or harassment. These systems, as Google points out, also embed social biases, and their output is frequently racist, or harmful in some other way.
This is mostly due to the way these systems are programmed. They're basically trained on massive volumes of data (in this case, many pairings of images and captions) that they study for trends and learn to duplicate. However, these models require a lot of data, and most researchers — even those working for well-funded corporate behemoths like Google — have decided that filtering this data thoroughly is too time-consuming. As a result of scraping massive amounts of data from the web, their models swallow (and learn to duplicate) all the vile vitriol you'd expect to see online.
"[T]he huge scale data requirements of text-to-image models [...] have prompted academics to rely largely on massive, mostly unrated, web-scraped datasets," Google's researchers write in their article. These datasets tend to represent societal prejudices, oppressive perspectives, and disparaging or otherwise detrimental connotations to minority identity groups, according to dataset audits."
In other words, in the whiz realm of AI, the well-worn saying of computer scientists still holds true: garbage in, garbage out.
The Image "encodes various social prejudices and stereotypes, including an overall bias towards creating images of people with lighter skin tones and a tendency for images showing different occupations to correspond with Western gender stereotypes," according to Google.
This is something that researchers discovered while testing DALL-E. If you ask DALL-E to create photos of a "flight attendant," for example, almost all the results will be female. When you ask for photographs of a "CEO," you'll get a bunch of white dudes, which isn't surprising.
As a result, OpenAI has decided not to release DALL-E to the public, although it does provide access to a small group of beta testers. In order to prevent the model from being used to generate racist, violent, or pictures, it additionally filters some text inputs. These safeguards go a long way toward limiting the technology's potentially negative applications, but the history of AI tells us that such text-to-image models will almost likely become public in the future, with all the worrisome implications that entail.
Image is "not fit for public usage at this time, according to Google, which aims to establish a new approach to measure "social and cultural bias in future work" and test future iterations. For the time being, we'll have to make do with the company's cheerful imagery, which includes raccoon royalty and cactus wearing sunglasses. But that's only the tip of the iceberg. If Image wishes to make an iceberg out of the unexpected effects of technological development, it can do so.
You must be logged in to post a comment.