Generative AI is a multimodal technology, with applications in text, image, video, audio, and code. Unfortunately, up until now, the actual usefulness of GAI in schools has been limited by technical and practical barriers.
ChatGPT, for example, is easy to access but problematic in the classroom due to its obscure terms and conditions and dubious privacy and data storage. There are also ethical concerns with its construction, the bias in the output, and the potential to generate inappropriate or misleading content.
Microsoft has attempted to address some of those issues with Bing chat, built on the same model but featuring more guardrails. As Microsoft rolls out its Copilot products, you can expect to see the use in classrooms increase.
But for image generation, we’ve had very limited possibilities in schools. Midjourney – arguably the most impressive of all the current models – lives on Discord servers. Discord is generally banned on school systems as a social network, and its also a largely unmoderated platform that isn’t recommended for use in education. It’s also not user friendly at all, and the user experience (UX) of Midjourney is pretty terrible.
OpenAI’s image generator, DALL-E, is… not great. By comparison to Midjourney, the current model produces weird, unrealistic images which occasionally veer off into total horror shows. Bing chat uses a version of DALL-E, so it can be accessed freely, but its usefulness is limited by its quality right now.
For comparison, here’s a photo of a group of people in Midjourney (left) and DALL-E via Bing chat. I’ve used “photo of a group of people” because current image generation particularly struggles with multiple individuals, as you can see from the facial distortions and other errors.
Use the slider to switch between “photos”.
We have open source models like Stable Diffusion, which can be accessed in a variety of ways such as via Dream Studio or apps built on community sites like Hugging Face. Again, the UX is clunky, and occasionally almost unusable. Open source image generators also have very few guardrails, and can be used to generate explicit content and deepfakes. Hardly something to throw at your year eight students.
All of the above options also have some serious ethical question marks over the scraping of images into the training datasets. Real artists’ styles can be easily emulated by typing their names, whether living or dead. There are several active lawsuits against companies like Stability AI and OpenAI for intellectual property violations.
Adobe’s Firefly has been around for a while now, but only recently released for full public use. It has several advantages over the above platforms for use in schools.
The model is trained exclusively on Adobe Stock images, public domain content, and images shared under open licenses. Nothing under copyright has gone into the training, unlike other models. Adobe does not use its customers content (photos and artwork stored on the Creative Cloud servers, for example) unless they are added to Adobe Stock.
Images generated with Firefly are tagged with metadata that identifies them as AI content, and they have worked on new licensing for AI generated images. This is an important step for transparency and avoiding the spread of deepfake or harmful images with GAI.
The technology has also been built into Photoshop, as well as a few extra features such as generative fill and generative expand. Many schools have Adobe Create Cloud licenses and/or Photoshop available for students, so Firefly is available through a simple email login with a school email address. In terms of privacy and student data, this means that students do not need to use a personal email account, and that logins can be centrally administered within the school.
Using Adobe Firefly
Firefly has another advantage: the user experience is much better than Midjourney, DALL-E, and Stable Diffusion models. I mentioned the simple login with school credentials (I use my Deakin University staff login). Once you’re in, you’re presented with this page:
Typing a prompt in natural language and hitting generate will take you directly to your first four options, and a few extra controls:
Using the panel on the right, you can then adjust the aspect ratio, content type (photo, graphic, art), and select from some pre-assigned “styles” such as digital art, steam punk, synthwave, watercolour, and so on. These appear as tabs under your original typed description:
Note that it is not possible with Firefly to generate art “in the style of X Artist’s Name”. If you try (even with deceased artists), you’ll get this error:
Firefly image quality
Quality is important – if you’re bringing image generation into the classroom, you want a reliable platform that isn’t going to generate useless, ridiculous, inappropriate or even horrific images. Like most image generation, Firefly does well in some areas, and not so well in others. This will be down to the frequency of images in the dataset, the quality of the dataset, guardrails to limit inappropriate content, and so on.
Here are a few examples of different styles and forms. For consistency, all of these are set to a 16:9 aspect ratio with no “style” tags.
Graphics, art, and digital art
Other Firefly features
As well as the text to image generation, Firefly offers a number of other modes and a few suggestions for features currently under exploration. At the moment, the following are available:
- Generative fill: background removal and replacement, “inpainting” to add new features to images
- Text effects: generative AI filled fonts allowing you to apply image over text
- Generative recolour: vector artwork recolouring
In the first example below, I’ve used generative fill to replace the background on the illustration of the leaf monster. In the next, I’ve added some coffee beans to the espresso photo. The remaining images show examples of text effects (recolour was unavailable at the time of writing).
Generative AI in Photoshop
For most purposes in the classroom, Firefly will probably suffice. For students using GAI in visual art, design, or media subjects, however, it might be worth exploring the features in Photoshop.
Photoshop has added generative fill, generative expand, AI assisted spot healing (similar to generative fill’s inpainting), and other features. The generative expand in particular has some great potential, allowing students to take an original image and expand well beyond the borders. Here’s a few examples with the coffee cup from earlier. Generative fill adds (or removes) coffee beans, generative expand takes us out to a shot of the table, then the entire cafe, and generative fill replaces the coffee with a vase of flowers.
Using Firefly in the classroom
In a future post, I’ll explore some ways to incorporate image generation into different subject areas, including English, Humanities, Visual Arts, and Design. For now, I’ll wrap up this post with a few observations on the images I’ve been generating in Firely and their usefulness in schools.
GAI could potentially replace stock images for students creating presentations, design folios, and other image-heavy tasks. In my experience students can fall down a rabbit hole of trying to find the perfect image in Google Image Search, and often end up using images which are copyrighted or not entirely suited to the task. As image generation improves, they’ll be able to get much more specific images.
Image generation could also be useful in imagining and designing settings, characters, products, and other ideas in a range of subjects. If a student can articulate exactly what they’re picturing, then they might be able to use GAI to generate some compelling visuals.
Honestly, I’m pretty impressed with what Adobe has put out there. I haven’t really used Adobe products a great deal since I was a VCE Media Study teacher, back around 2012. Photoshop has come. along way since then, and I’m sure that both Firefly and PS’s generative features will make image generation in the classroom a lot more accessible.
Have any questions or comments about this post, or interested in professional development for GAI? Get in touch: