JBurton
Signtologist
I thought, maybe I'm being stupid, so I waited until the end of the video, clicked on an image and the video started again... this is all I get.
-
I want to thank all the members that have upgraded your accounts. I truly appreciate your support of the site monetarily. Supporting the site keeps this site up and running as a lot of work daily goes on behind the scenes. Click to Support Signs101 ...
Photoshop Beta / Generative Fill
- Thread starter jochwat
- Start date
Fred Weiss
Merchant Member
You should be able to get to a gallery of images where you can click on any image to access the prompt. You can then change the prompt text to get different results. Here's an example I just did changing the prompt from New York City to Miami Beach.
It's... a weird program. It has a long way to go, but it does show that theyre getting better with AI art / prompts.
I I downloaded a random stock image of people, I wanted to see how it'd do selecting a person, typing no prompt and doing a generative fill.
I first selected the guy on the left - whole body... Hit fill, which they say is supposed to remove objects and people.... but nope, it changed him to a different black guy - I selected just the head to the guy on the right... and it did the same thing. So I selected the girls head...and it did the same thing.
It blends it pretty good, actually curves the neck / suits so it doesnt look "too" fake..
I just noticed I can scroll through 3 options for each - It also changed the guy on the left into an asian girl twice, the other head images dont look good and look deformed.
-Just hilighted the guys shirt, and told it to change it to a T-shirt.
Change to swimsuit... obviously didnt go so well (Note: I'm not a perv 
Just didn't want to do another T-shirt change since those are easier, and just changing her face left her arms white, so I wanted to see what itd do)
In 5 mins of playing I've got it to change quite a few things.... Kind of "cool", but clearly still a work in progress. It looks like you can actually tell it to do stuff, VS just adding stuff into photographs...
It does some weird crap, and has a long way to go before people need to worry about their jobs though. I think it'll benefit graphics artsists to do quick / easy changes... but nothing to fear yet.
I I downloaded a random stock image of people, I wanted to see how it'd do selecting a person, typing no prompt and doing a generative fill.
I first selected the guy on the left - whole body... Hit fill, which they say is supposed to remove objects and people.... but nope, it changed him to a different black guy - I selected just the head to the guy on the right... and it did the same thing. So I selected the girls head...and it did the same thing.
It blends it pretty good, actually curves the neck / suits so it doesnt look "too" fake..
I just noticed I can scroll through 3 options for each - It also changed the guy on the left into an asian girl twice, the other head images dont look good and look deformed.
Just didn't want to do another T-shirt change since those are easier, and just changing her face left her arms white, so I wanted to see what itd do)In 5 mins of playing I've got it to change quite a few things.... Kind of "cool", but clearly still a work in progress. It looks like you can actually tell it to do stuff, VS just adding stuff into photographs...
It does some weird crap, and has a long way to go before people need to worry about their jobs though. I think it'll benefit graphics artsists to do quick / easy changes... but nothing to fear yet.
The thing with AI as well, every software is going to have its own terms of service that restricts it... making it pretty much useless. Yes, there needs to be rules... and maybe photoshop is just being overly sensitive right now.
I tried to add "Fighter jets" to the vancouver skyline... but couldnt because it was against TOS.
I tried to add "A horde of zombies chasing people" to a different image...against TOS. Tried just zombies... nope, against TOS. Werewolves were perfectly fine though....
Everytime I add a bus or something, it has foreign writing on it... I'm guessing photoshop trained their models on foreign stuff to try to avoid the copyright debacle.
I wish they focused more on AI'ing the software to do basic tasks automatically, than on generating images. It'd be nice to be able to select someones shirt, tell it to change it to red, and have it do a halfway decent job - It'd be a lot more useful than what its trying to do now
I tried to add "Fighter jets" to the vancouver skyline... but couldnt because it was against TOS.
I tried to add "A horde of zombies chasing people" to a different image...against TOS. Tried just zombies... nope, against TOS. Werewolves were perfectly fine though....
Everytime I add a bus or something, it has foreign writing on it... I'm guessing photoshop trained their models on foreign stuff to try to avoid the copyright debacle.
I wish they focused more on AI'ing the software to do basic tasks automatically, than on generating images. It'd be nice to be able to select someones shirt, tell it to change it to red, and have it do a halfway decent job - It'd be a lot more useful than what its trying to do now
Fred Weiss
Merchant Member
I think what they're up against is about a dozen already released or near release competitors. While they may have some things according to their rules and codes, their choices also have to end up in a product that will earn a profit.
fred,
isn't the interesting distinction here that they are not SEARCHING an existing pool of images, they are CREATING these images from whole cloth.
if you follow what is happening in hollywood now, they want to use these technologies to produce the 100 + language variations you might release a movie in. this eliminates a VERY PROFITABLE SECTOR for actors, dubbing the audio for interntional release in studio, post production
i believe it was 1999.when david chase/hbo required actors to do the audio for each scene in english/spanish/ french AS PART OF original production costs. their product ws redy for immediate international release, without the months long dubbing process.
current contract negotiations involve making your voice available for AI post production use
White Haus
Not a Newbie
I'll need to play around with it some more. Tried using it on one of our recent wedding picture (Hands showing rings) and it said it violated some kind of policy. Huh?
Definitely pretty interesting technology. Will explore more. Love the examples you guys have shared already lol.... we should have a contest for the best submission.
Definitely pretty interesting technology. Will explore more. Love the examples you guys have shared already lol.... we should have a contest for the best submission.
jochwat
Graphics Department
As ikarasu mentions, there does indeed seem to be a lot of TOS conflict. The aforementioned zombies, as well as missiles (I changed it to rockets), as well as "kids in swimsuits playing in the water" which I understand, but when I put "family playing in the water", well, there were the kids in swimsuits, along with their parents / guardians. So yeah, a long way to go, but it sure does seem to be ramping up. In the "other stuff we're dreaming up" section is a "Text to Vector" system that might spook me a bit.
jochwat
Graphics Department
The things you were adding to your image are hilariously identical to what I was adding while playing around with this at first... other than Sasquatch. I did go with the obligatory dinosaurs, cows, and the occasional forklift.
JBurton
Signtologist
Ok, totally irrelevant to the higher level discussion, but it sounds like you 'misused' the selection tool? Like, selecting the guy on the left, ps took that as the reference for your command 'fill', so it began filling the image with black folks? IDK, just seems plausible.
Thanks for the series of images and play by play!
Now if you'd used beefsquatch, it wouldn't have been confused, but ya chose something with no real world reference, so you get... idk what that is.
Fred Weiss
Merchant Member
Not sure where you got that information. As an Adobe Stock contributor I can tell you that Adobe is busy acquiring rights to existing images including mine.
i have been told this by a couple computer consultant friends.
i thought it was reinforced when the presenter said "you will not get the same car, but a similar one"
Notarealsignguy
Arial - it's almost helvetica
Does anyone remember smarter child, the chatbot? It was fun to ask it dirty questions and see what it'd say but was relatively useless. This seems like the same thing.
I have no doubt what it did was analyze the image, then decide to add someone else based on there being 2 others in the image... but according to photoshops own how to - Which is exactly what I did... selected a person, hit generate with no prompt...And it just completely replaced him instead of removing him, then proceeded to change everyone elses skin color when I selected them without a prompt as well. 2 Of the people were full selections, then when I saw what it was doing I started to only select heads to see if it would still do it.
To be clear... you get 3 options, and If I went to option 2/3 it changed to an asian girl for option 2, and a white person for option 3- Option 2/3 for the girl was a completely garbled face.
I'm more surprised that rather than removing objects, it decided the smart thing to do was to put a random person in his place!
[Edit] And I do get its a beta product. But you'd think theyd default, or prioritize removing an object if nothing is prompted - if I wanted it to replace the guy, I could have prompted it with "Replace with someone else" or something. Still a neat product, it just has a long way to go, feels rushed compared to Dalle / Midjourny.
Last edited:
Boudica
Back to "educational purposes"
We used to do that with Alexa. She has some pretty good fart jokes.
Very cool. So far I've used it to add bleed to a graphic wall print and it did an excellent job.
One gotcha: it generates a maximum size of 1024px x 1024px. If you select and generate an area with a side greater than 1024px, it will generate the largest side at 1024px and scale up, which could lead to some terribly pixelated results.
The workaround is to generate in portions smaller than 1024px on the largest side. Worked great for adding bleed. Here's an example of how to do that:
One gotcha: it generates a maximum size of 1024px x 1024px. If you select and generate an area with a side greater than 1024px, it will generate the largest side at 1024px and scale up, which could lead to some terribly pixelated results.
The workaround is to generate in portions smaller than 1024px on the largest side. Worked great for adding bleed. Here's an example of how to do that: