Based on the location, the image variations are stuck at certain stereotypes

[mellonin]

Hello, 

I tried Firefly's text to image adaptation. I wrote "two people talking" and it showed up two Arabic women/men talking in traditional Arabic clothes with a mosque background. I changed the title to "two modern people talking" and the similar images appeared. Is it possible that because my location shows up as Eastern Europe/Middle East, the images appear to the location accordingly? If so, it is so stereotypical and a weird bug. Please consider fixing this problem, it is really weird to get this outcome even with detailed explanations. For example, I tried "two modern people talking in Izmir" -Izmir is a city in Turkey- and it showed up only two Arabic men again, no women whatsoever, with traditional clothes as if I wrote in Saudi Arabia... I just don't understand why this happens. Thank you for considering the solution.

[Jill_C]

I think you're correct in that Firefly assumes you want people to look like those in your part of the world. I'm in California, USA and when I requested "two people talking" I got 2 vaguely Hispanic-looking dark-haired people. When I asked for "two modern people talking", I got slightly younger Hispanic people. You have to be much more specific in terms of the appearance of the people. I requested "Two blonde American people talking, one man, one woman" and got young couples that look like Caucasian college students. 

[droopydog500]

It is not just people, but also objects (like flags) that are influenced by where it thinks you are located, as is referenced in this post: P: Union Jack appearing in images - Adobe Community - 14211823

As Jill said, it is using where it thinks you are located because it thinks you will want people and things that appear to be from where you are. I would be nice to be able to turn that behaviour off. The best work around now is to specify regions, locations, races, or cultures in your prompts. 

[Jill_C]

Frankly, if I lived elsewhere (other than USA, which we know is a huge melting pot anyway), such as Asia or Africa, I would be annoyed if I requested images of , say, "two college aged students walking and holding hands", and I always got blonde-haired white people!

[Zaid Al Hilali]

Interesting. I'm located in Dubai (Middle East) and here is what I got…

[Jill_C]

Under the content section on the right side of the Firefly screen, be sure to select "Photo" rather than "Art" to avoid getting these cartoonish looking characters.

[Jill_C]

Let's compare results. Run this prompt "two young people in Dubai walking and holding hands, one woman, one man" and upload your results. Mine are attached. I'm in California, USA.

[Zaid Al Hilali]

Kind of similar results to yours, Jill

[Jill_C]

Interesting. Is it typical for young women to be so modestly dressed? When I initially asked for "middle eastern" people, all of the women were wearing a hijab.

[Zaid Al Hilali]

So true. Apparently we need to mention hijab in the prompt.

[Jill_C]

I could not get rid of the hijab, not matter what wording I tried: "No hijab", "without hijab", "no headcovering", "no hat".... Those hijabs are stuck on there to stay !

[Zaid Al Hilali]

Jill, you may have answered the first poster. If part of your description is now stuck with every new generate, so maybe @mellonin typed a prompt sometime earlier that includes a "mosque" or "Muslim" which is affecting every new generate!

[Jill_C]

I didn't use those terms in any of my prompts, but perhaps Firefly includes some of the details from earlier prompts?

[droopydog500]

I believe there is some base information carried from generation to generation in the same thread (my terminology). Try going back to the main firefly screen and start a new session.

[Jill_C]

You may be right; I logged out entirely and logged back into my Adobe account, and the hijabs have gone away...

[Zaid Al Hilali]

Now based on these findings, I wonder if Adobe have documented these in their FAQs?

[nmycps]

This issue of ethnicity and your geolocation determine to a degree the behind-the-scenes-generated output in Firefly, and this could use more attention. While it isn't difficult to specify ethnicity in people or objects in the scene, it can be counterproductive to have to write prompts to counteract automatic geolocation settings. The prompt is still the main rudder of course in determining output, but to propose building in some sort of 'ethnicity strength' slider that will strengthen your prompts that need to be 'culturally accurate', or to lessen the strength so that it includes more randomness in generating people from different ethnicities.