So, while at the ELT Freelancers Awayday in Oxford last week, AI was a big topic that was talked about. One service that intrigued me was Adobe Firefly. Firefly is a service of generative AI for images and is currently in beta testing. So I thought I'd test it out.
I don't know about you, but the majority of image-generating AI services I've seen tend to produce weird, distorted faces - and there are still plenty of those in Firefly, but some that aren't too bad. If you can see the image at the top, this was created with Firefly. The prompt was 'Two men getting married' and this was probably the best of the four it came up with. What I like about this, is that I didn't specify ethnicity, but what I don't know is that the dress is authentic or not. Would need to do some research on this.
If you've followed me for a while, or know the kind of work I do, then you'd know I'm interested in how LGBTQIA+ identities could be represented. One thing I've found out with Firefly is that it needs to be trained on what the trans flag is, as well as the pride progress flag.
Now, the prompt for the above image was 'a teenager's messy bedroom with the trans flag on the wall'. That is not the trans flag on the wall, and when I changed the prompt to the pride progress flag, it didn't know that either. Apart from that, the face in this one is rather distorted, but I like the mess that it created.
Moving on, I wanted to look at how Firefly represented people. I typed in a prompt for a trans woman. I didn't, and don't, want to post what it came up with because I thought it might be offensive. Needless to say, the AI needs better training on this. However, my next prompt was for a trans man in a kitchen, and it came up with this:
This could well be a trans man (who is indeed a man, by the way) in a kitchen. But there's no authenticity there. Of course, we can't (and shouldn't) judge who people are by appearances, so maybe this image is OK. What I like about this image is the body shape - fairly good in terms of representation.
I then wanted to see what Firefly would bring for a portrait of a gay male couple with a disability and one for a lesbian couple in a living room.
Again, there are issues with both images, and with these ones I specified the ethnicity.
However, I'm encouraged with the level of quality of these images, even with the most basic of prompts. It's clear that in order to use these services well, creating prompts is a skill that we need to learn and hone in order to get the best images.
I realise that this is just a very, very basic test of Adobe Firefly, but I wanted to share what I'd found out and I'd be interested to see what other people come up with, and what aspects of image creation AI needs to improve on.
If you'd like to have a go yourself, there are free credits on the Adobe Firefly website, which is what I used. You can use this link here - https://www.adobe.com/uk/sensei/generative-ai/firefly.html
Let me know how you get on!