INPUT: TEXT, OUTPUT: IMAGES - AI TEXT-TO-IMAGE GENERATORS

     Text-to-image generators use the input of words to generate the output of an image. The words input can be as simple or specific/complex as desired. As discussed throughout this site over the summer, most of the images generated for the site came from the input of the words "a perfect world," as well as other words modified by "perfect," such as "a perfect tree," "a perfect smile," etc. Text-to-image generators, ChatGPT and similar systems fall under the category of "generative AI." In an article for MIT News, author Adam Zewe defined generative AI as "a machine-learning model that is trained to create new data, rather than making a prediction about a specific dataset. A generative AI system is one that learns to generate more objects that look like the data it was trained on." (n6, click here to read the full article explaining generative AI.) It is probably important to remind the reader here that there is no one single AI system. Each content-generating system, whether for image or text generation, is owned and or operated by a company or organization, is guided by proprietary algorithms, is trained on different but massive datasets, and (in most cases) offers different options to the end user. At this time, all text-to-image generators must be used online, because the generative function requires massive computing power housed at AI data centers. Though probably the subject for a separate essay, the amount of power required to operate these AI data centers is growing exponentially. Though forecasts vary, some project that "data centers could account for 25% of U.S. power consumption by 2030 based on surging demand for AI chatbots and AI training." (n7)

Getting Started with a Text-to-Image Generator

      There are a number of text-to-image generators available, and three in particular will be discussed below. Plans and access vary from free and free trials to different levels of subscriptions. Since these are constantly changing, it is recommended that the user check for the most current information for each image-generating system. For most systems, the words input to generate an image are referred to as prompts, and most generators allow for a certain number of credits, tokens or some other measure of the number of images able to be created in a given time period. Once word prompts have been input, a fairly common output generated is four square images based on those prompts. Some systems also allow for changes to the size of the images which are generated.

Microsoft Bing Image Creator login screen image.

Microsoft Bing Image Creator. The Microsoft Bing Image Creator can be accessed at www.bing.com/create or www.bing.com/images/create, and the log-in screen is shown in the picture above. This image generator is powered by Open AI's DALL-E 3, and it is free to use with a Microsoft account (if you use Hotmail, for example, you can log in using the same email address and password). If you click on the words "Learn More," found just under the bright pink "Join and Create" box, you can find out a bit more about the system and suggestions for structuring the prompts. Also, if you are logging in on a desktop computer, once you are logged in you will see a wide variety square images like the ones see above. Holding a cursor over the image will reveal the prompts entered to generate it.

Imagine AI Art Generator. The Imagine AI Art Generator can be accessed at www.imagine.art, and it is a product from the company Vyro (www.vyro.ai). Vyro describes Imagine AI as "an innovative art generator accessible on the web, Android and iOS platforms, designed to provide individuals with easy access to the latest generative art technologies." Though there is some free use of Imagine AI permitted, there is no free trial, and full use of most premium features requires the purchase of a weekly or monthly plan. Plan costs can be viewed on the Imagine.Art website. In addition to text-to-image generation, Imagine AI offers applications including AI Video, Ideate (turning doodles to art), and Image Remix (altering exiting images), plus "mini-apps" Realtime Generation, Creative Upscaler and Character Consistency. Different sizes for generated images (other than square) also can be specified. As stated in various locations on this site, most of the images used this summer were generated by the Imagine AI Art Generator.

Midjourney. The above image shows a login screen for Midjourmey found at https://discord.com/invite/midjourney. You have to create a Discord account in order to use Midjourney, and according to the information on the screen shown above, there are more than 20 million Midjourney members. At www.midjourney.com Midjourney is described as "an independent research lab exploring new mediums of thought and expanding the imaginative powers of the human species." You also can sign up and log in from the Midjourney.com site. Beginning users might want a bit of direction in using Midjourney, and a good place to start is at https://docs.midjourney.com. On that page you can find a Quick Start Guide, links to "Getting Started" and "Using Discord," and access to articles on subscription plans, terms of service,prompts and style reference. The creators of Midjourney note at https://docs.midjourney.com/ docs/stealth that Midjourney is an open-by-default community, and all image generations are visible at Midjourney.com. However, subscribers to the two highest plan levels can use Stealth mode which prevents your images from being visible to others on the Midjourney website. Read more on the web page listed.

     On this site, you can find images created in Midjourney by Guest Photographer Ann Elliott Cutting on the All Cities By City concert page (click the link to view the images in a new window). She says of the image generator that "Midjourney has consistently stood out to me for its refined style. I also use Stable Diffusion, but I find that Midjourney's ability to incorporate style references is particularly helpful. I've been using Midjourney since Version 2, and now, with the release of Version 6.1, it's been about a year of exploring its evolution." She also says she is "fascinated by the unique qualities of these early AI creations," and she can see a future in which photography may "evolve into a hybrid process - combining real elements like people, clothing and products with AI-driven enhancements."

     Information on a couple of other AI-related programs may be of use to those interested in AI-generated content. Most text-to-image AI-generated images are relatively small/low resolution. Those wishing to enlarge the images will likely need a program for "upscaling" the images. In the text above regarding "Theatre D'Opera Spatial," its creator referenced the program Gigapixel AI as the means he used to upscale the image, increasing its resolution and size. Information on Gigapixel can be found at www.topazlabs.com/gigapixel.

     Those who use Adobe programs like Photoshop or Illustrator or who are members of the Adobe Creative Cloud are likely to have received emails over the last couple of months advertising the new AI capabilities to be incorporated into Adobe products. The image here on the left is from an email advertisement for the new Generative Fill function in Photoshop. An example of how the Generative Fill function in Photoshop will work (derived from the gif image pictured in the ad) is included below. As with other AI systems, most generative AI functions at this time will need to be used online due to the computer power necessary to back those functions. Adobe has named their AI Firefly. Details about how Firefly AI will be incorporated into various Adobe products can be found at www.adobe.com/ products/firefly.html.

     Earlier this month, an email message under the subject heading "September Community Newsletter: Generative AI" went out to Adobe Creative Cloud customers from David Wadhwani, President of Digital Media at Adobe. In the letter addressing Adobe's approach to AI, he said the following:

     "'Adobe has always looked for innovative ways to implement new technologies that help creators realize their vision. For the past few years, we've been very focused on delivering innovation that we believe will make the communities we serve more productive. Much of this innovation involves harnessing AI breakthroughs to reduce the tedious and repetitive work that often makes up a significant portion of the creative process. This work, and generative AI advances in general, have caused some in the creative industry to ask some fundamental questions about the definition of creativity, about ways to protect creator rights and even about the future of creative industries. . .

     First and foremost, I want to be clear that we believe AI can be a tool for, and not a replacement of, human creativity. That's why our approach toward AI is to build technology that elevates your creative workflows and integrates directly into the tools you already use . . . Second, we believe generative AI can and should be developed responsibly starting with respect for the rights of creators. That's why we only train Adobe Firefly, our family of generative AI models, on content where we have license or permission, and we compensate Adobe Stock contributors for use of their content to train Firefly . . .

     . . . We've created a webpage where we clearly articulate our approach to generative AI with Adobe Firefly, , , In short, we do not and never have trained Firefly on your content in Creative Cloud, and we don't claim any ownership of your content - including content you create with Firefly. We only train Firefly on content where we have explicit rights, we don't mine content from the web and we compensate contributors through Adobe Stock. . .

     . . .We've been working across the industry and with govenments worldwide to protect the creative community and your ability to thrive. As part of this effort, we co-founded the Conent Authenticity Initiative in 2019, which is now supported by over 3,300 members and aims to provide transparency on how content was generated and who created it. We're also actively helping to shape and advocate for policies and legislation like the FAIR Act, which aims to help protect artists from the harmful and unfair use of AI to replicate their work." (n8)

     The process of adding a lighthouse to a starting image in Adobe Photoshop using the new Generative Fill function is depicted below, and the Content Authenticity Initiave mentioned will be discussed a little later in the text. First, however, the U.S. Copyright Office's treatment of AI content and the Office's current actions regarding AI-generated materials will be outlined in greater depth.

Adding a lighthouse with Adobe Photoshop Generative Fill: 1. begin with base image, 2. open Generative Fill function, 3. enter word prompt in box and click "generate," and 4. select lighthouse from generated images.

AI-GENERATED MATERIALS AND U.S. COPYRIGHT OFFICE POLICY

     In addition to the U.S. Copyright Office case listed above for the image "Theatre D'Opera Spatial," another case involving the attempted copyright of Midjourney-generated AI images has received considerable attention. This second case involved the copyright for a comic book titled "Zarya of the Dawn." Initially, based on a copyright application for the work which neither disclosed AI-generated content nor disclaimed any part of the work generated by Midjourney, a copyright was granted. However, as outlined in correspondence between the U.S. Copyright Office and the attorneys for the comic book's author, Kristina Kashtanova, "Shortly after registering the Work, the Office became aware of statements on social media attributed to Ms. Kashtanova that she had created the comic book using Midjorney artificial intelligence. Because the application had not disclosed the use of artificial intelligence, the Office determined that the application was incorrect, or at a minimum, substantively incomplete." (n9) After additional review and correspondence, the original certificate was cancelled and a new, more limited registration was granted. That new certificate of registration covered "the original authorship that Ms. Kashtanova contributed to the work, namely the 'text' and the 'selection, coordination and arrangement of artwork created by artificial intelligence and text created by the author', . . ." but specifically excluded "artwork generated by artificial intelligence." (n10) The correspondence from the U.S. Copyright Office concluded by saying "the Office will add the following annotation to the new certificate: 'Reason for Reregistration: VAU001480196 cancelled pursuant to 37 CFR 201.7 (c)(4) for failure to exclude non-human authorship.'" (n11)

     Two other significant points were contained with in the "Zarya" correpondence: the human authorship requiement for the granting of a copyright, and the description of how the office views AI-generated content, in this particular case, content created by Midjourney.

Human Authorship Requirement. One of the fundamental principles of U.S. copyright law, practice and precedent is the requirement for human authorship. "The U.S. Copyright Office will register an original work of authorship only if the work was created by a human being. U.S. Copyright Office Compendium of U.S. Copyright Office Practices 306 (3d ed. 2021). The copyright law only protects "the fruits of intellectual labor" that "are founded in the creative powers of the mind." Trade-Mark Cases, 100 U.S. 82, 94 (1879). Because copyright law is limited to "original intellectual conceptions of the author," the Office will refuse to register a claim if it determines that a human being did not create the work. . . Consistent with the law, the Office will not knowingly register works produced by a machine or mere mechanical process that operates randomly or automatically without sufficient creative input or intervention from a human author." (n12) It appears that in this case as well as others, claimants have attempted to argue, among other things, that extensive input of prompts (both initial and image-refining) into an AI text-to-image generator qualifies as the sufficient creative input and intervention of a human author. That argumemt has not been accepted by the Office, as detailed below.

Copyright Office View of AI-Generated Content in Zarya Case. The U.S. Copyright Office correspondence in the Zarya case gives a very detailed explanation of how AI-created content, in this case images, are viewed. Because it makes a pointed comparison to works made for hire by actual human artists, relevant details from the text are included below.