With the "Deus Ex Mach­ina? — Test­ing AI Tools"-series we want to show you dif­fer­ent tools, that aim to sim­pli­fy writ­ing, design and research by using Arti­fi­cial Intel­li­gence. More on the "Deus Ex Machina?"-series can be found here.

A con­tri­bu­tion by Helena Bodem (study pro­gram Sci­ence-Media-Com­mu­nic­a­tion, KIT Karlsruhe)

Overview

The digit­al tool Play­ground AI (playgroundai.com) is an image gen­er­at­or that provides an easy way for both non-experts and exper­i­enced users to cre­ate and edit images. At the heart of the tool is a "text to image" AI that allows users to gen­er­ate images or modi­fy exist­ing images using spe­cif­ic text instruc­tions (prompts). Because a "text to image" gen­er­a­tion leaves a lot of room for ran­dom res­ults, anoth­er text input is pos­sible below the prompt field: the so-called "neg­at­ive prompts", which can be used to expli­citly exclude cer­tain fea­tures or ele­ments from the image generation.

A range of pos­sible fil­ters and con­trols can be used to fine-tune the image gen­er­a­tion. How­ever, since the under­ly­ing effects are not intu­it­ively access­ible, it is advis­able, espe­cially for non-pro­fes­sion­als, to stay with­in the stand­ard set­tings for the time being. Espe­cially at the begin­ning, users may be over­whelmed by the vari­ety of res­ults: even if you simply leave the same text prompt and click on "Gen­er­ate" sev­er­al times, the tool pro­duces dif­fer­ent image inter­pret­a­tions. Users can also upload their own pho­tos and modi­fy them with the AI. Both an exist­ing and a gen­er­ated image can be edited by mask­ing parts of the image and chan­ging masked objects or adding new objects via "Edit Instruction".

Any­one who wants to use the tool needs a Google account. In the free basic ver­sion, up to 1000 images per day can be gen­er­ated or edited. The images gen­er­ated in this way can not only be down­loaded, but also scaled up in qual­ity. While Play­ground AI allows users to use the images for private and com­mer­cial pur­poses, copy­right and eth­ic­al issues remain with this image gen­er­at­or as well.

The AI behind the application

The tech­no­logy that powers Play­ground AI is based on vari­ous ver­sions of the Stable Dif­fu­sion AI mod­el. Its name is derived from the com­pany name Sta­bil­ity AI, which was instru­ment­al in the devel­op­ment of Stable Dif­fu­sion. Sta­bil­ity AI inves­ted sig­ni­fic­ant amounts of GPU com­put­ing power to enable the train­ing of this mod­el. Although there are vari­ous "text to image" AI mod­els, with Dall‑E prob­ably being the most prom­in­ent, Stable Dif­fu­sion stands out in that it is the only one of its kind to have been pub­lished in open source format: The mod­el is com­pletely open to pub­lic scru­tiny and can be mod­i­fied by the developer community.

In its mode of oper­a­tion, Stable Dif­fu­sion uses the prin­ciple of dif­fu­sion, whereby images are mod­i­fied by repeatedly adding Gaus­si­an noise, i.e., the pixels of an image are blurred until it con­sists only of dif­fer­ent col­oured spots or dots. The AI mod­el is then trained to recon­struct the "noisy" image, cre­at­ing a vari­ation of the ori­gin­al image.

In order for "text to image" AIs to gen­er­ate the most diverse images pos­sible, a large set of train­ing data is needed. In the case of Stable Dif­fu­sion, a data set known as LAION 5B was used. It con­tains an impress­ive 5.8 bil­lion HTML links to cor­res­pond­ing image-text pairs. How­ever, in cre­at­ing this set, little con­sid­er­a­tion was giv­en to con­tent rat­ing, copy­right or pri­vacy of the images. For example, one artist dis­covered med­ic­al images in the set that had been taken of her before and after an oper­a­tion. The artist now has little chance of hav­ing these images deleted. Accord­ing to LAION's terms and con­di­tions, own­ers of images can only request the dele­tion of links from the set if iden­ti­fi­able data can be found on the linked images, e.g., name, tele­phone num­ber or address.

(Fur­ther inform­a­tion on eth­ic­al and leg­al con­flicts can be found in the sec­tion "Usage in sci­ence com­mu­nic­a­tion" and in the info box).

The rhetorical potential of the tool

People with cre­at­ive and design back­grounds are likely to be par­tic­u­larly attrac­ted to Play­ground AI. How­ever, the tool's abil­ity to quickly visu­al­ize ideas and make adjust­ments to images without pri­or know­ledge of pro­fes­sion­al edit­ing pro­grams also makes it inter­est­ing for a broad­er audience.

How­ever, it should be emphas­ised that Play­ground AI is not com­pletely intu­it­ive to use. For­mu­lat­ing an effect­ive text prompt requires prac­tice. For effect­ive prompt­ing, it is advis­able to use "tags", short image descrip­tions that usu­ally con­sist of one or two words and are strung togeth­er sep­ar­ated by semi­colons. There are also a vari­ety of fil­ters, adjust­ment para­met­ers and edit­ing options to help refine the res­ults. This can be over­whelm­ing for begin­ners at first.

A char­ac­ter­ist­ic fea­ture of "text to image" mod­els is a high ran­dom factor when gen­er­at­ing images. This is par­tic­u­larly evid­ent when using Play­ground AI: users can press the "Gen­er­ate" but­ton repeatedly without chan­ging the set­tings or prompt, and still get dif­fer­ent res­ults each time. In fact, the res­ults are so var­ied that one quickly wants to move on to vary­ing a par­tic­u­larly inter­est­ing image. Cre­at­ing vari­ations of a par­tic­u­lar image is pos­sible, for example, via the gen­er­a­tion num­ber of an image, the so-called seed. (In the user inter­face "Board" there is even a select­able action for this).

Anoth­er annoy­ance is that the tool seems to have dif­fi­culties with the gen­er­a­tion of spe­cial details such as fin­gers and toes. How­ever, to fix this and sim­il­ar issues, Play­ground AI provides some lengthy You­Tube tutorials.

The tool has two user inter­faces — Board and Can­vas — which are quite sim­il­ar in func­tion­al­ity but have dif­fer­ent emphases for use. While Can­vas is par­tic­u­larly suit­able for gen­er­at­ing images in quick suc­ces­sion, Board is optim­al for cre­at­ing vari­ations of an already cre­ated image.

An import­ant note about the board: here, images are auto­mat­ic­ally shared with the com­munity unless the share func­tion is dis­abled before gen­er­a­tion. This set­ting must be re-set before each use.

Usage in science communication

Dur­ing a self-test the pos­sib­il­it­ies and lim­it­a­tions of the tool became evid­ent. Play­ground AI proved use­ful for design­ing a cov­er image for a journ­al­ist­ic art­icle on arti­fi­cial intel­li­gence. A simple prompt like "Arti­fi­cial Intel­li­gence" com­bined with vari­ous fil­ters and set­tings quickly deliv­ers innov­at­ive image con­cepts. With adjus­ted prompt input and cor­rect image ratio, an appeal­ing res­ult can be achieved.

In con­trast, the attempt to cre­ate a simple dia­gram with four labels failed. Although there are prob­ably ways to give the tool more pre­cise instruc­tions, the high ran­dom­iz­a­tion factor in image gen­er­a­tion by Play­ground AI still stands in the way of cre­at­ing accur­ate dia­grams. And what about sci­entif­ic rep­res­ent­a­tions? For example, is it pos­sible to gen­er­ate a micro­scop­ic view of the bird flu vir­us? Even though the vir­us has already been extens­ively researched, the pro­ject proved dif­fi­cult without a con­crete image template.

Of course, sci­ence com­mu­nic­a­tion does not only take place via art­icles and sci­entif­ic papers. Social media plat­forms are also import­ant dis­tri­bu­tion tools. Here, col­or­ful images are often needed to attract the atten­tion of users. Espe­cially for this pur­pose, the tool quickly deliv­ers prom­ising results.

Good to know: Since the AI mod­el Stable Dif­fu­sion was trained with images of 512 x 512, Play­ground AI gen­er­ates par­tic­u­larly good images in this aspect ratio. For image series, it is there­fore advis­able to leave the pre­set height and length at 512 and to sub­sequently upscale selec­ted images (right-click > "Upscale by 4x").

If you are now hold­ing your self-gen­er­ated images in your hands, you are prob­ably ask­ing your­self ques­tions about image rights and use. In the FAQs (as of 04.09.2023), Play­ground AI gives a gen­er­ous answer to the ques­tion "How does copy­right work? Do I own the images I cre­ate?": All rights, titles and interests in the images cre­ated are left to the users — both for private and com­mer­cial purposes.

How­ever, the mat­ter is presen­ted in a more dif­fer­en­ti­ated way in the terms of use. Here, Play­ground AI, as well as its leg­al suc­cessors and agents, reserves far-reach­ing rights of use. This extends not only to con­tent gen­er­ated with the tool, but also to any images con­trib­uted by users. Fur­ther­more, those who fail to cre­ate their images in a private mode (which must be done again at the begin­ning of each board ses­sion) also grant oth­er mem­bers of the plat­form a non-exclus­ive licence to use them.

What this means in con­crete terms: Any­one who feeds their own image into the tool, e.g. as a tem­plate or for edit­ing, should always be sure that they own the cor­res­pond­ing rights to it. Because you have to reck­on with the fact that Play­ground AI will con­tin­ue to use this image. In addi­tion, Play­ground AI makes it clear in its terms of use that users them­selves are respons­ible for their activ­it­ies and points out that gen­er­ated images may unin­ten­tion­ally resemble pro­tec­ted images. So in prac­tice, it is prob­ably advis­able to at least do a Google Image search before using a gen­er­ated image for pub­lic­a­tion purposes.

It should also be borne in mind that, as things stand, it is still very uncer­tain wheth­er copy­right or copy­rights can be claimed for AI-gen­er­ated images. In the US, courts have already rejec­ted copy­right claims. So while com­mer­cial use is unlikely to be a prob­lem, it may be dif­fi­cult to take leg­al action if anoth­er per­son uses one's AI-gen­er­ated work.

Wrap-Up

"Play­ground AI" deliv­ers what its name prom­ises: it offers a cre­at­ive play­ground where users can try their hand at words and set­tings to their heart's con­tent to quickly get visu­ally appeal­ing res­ults. But the ini­tial euphor­ia can quickly turn into frus­tra­tion if the desired res­ult seems unat­tain­able. Par­tic­u­larly in the con­text of sci­ence com­mu­nic­a­tion, where accur­acy and pre­ci­sion are para­mount, Play­ground AI has its lim­it­a­tions and is more suit­able for cre­at­ing dec­or­at­ive accessories.

Fur­ther research on the tool evokes mixed feel­ings: On the one hand, it is motiv­at­ing to push Play­ground AI to get even bet­ter res­ults through improved prompts and more pre­cise set­tings. On the oth­er hand, one quickly comes across inform­a­tion that images were used for the tool's AI whose cre­at­ors neither agreed to nor were paid for them to be used.