I'm finding the process of using the generative fill to be a lot like using a slot machine.
The results tend to be almost entirely out of my control, and ultimately on most occasions I will settle for some barely passable output as a winner rather than what I really wanted, even though none of what I'm prompting for is really all that complicated as an image, or so it seems.
I mentioned in a prior post that I could not get, after hundreds of tries, the right color shirt on the subject of a photo. Since then, I have tried over and over, and it is just a matter of random luck. Ask for a white teeshirt, and I get something else so often that it almost seems to be wilfully disobeying. Prompt for shorts and get longs. Ask for longs and get legs twisted up like pretzels. Try to reposition an arm by removing and replacing, and it just won't go away.
Prompt for a fist or an open hand, and just get question marks, as if these things do not exist in our world.
Prompt for a shirt with vertical stripes and get horizontal stripes 98% of the time.
Ditto for asking for a particular pose -- especially in the text2image -- and you get that only occasionally. Ask for the single subject to be standing and facing the left of the image, and that comes up once in a while, but only after pretty much after every other direction has shown up, including back of the head to the viewer. It is not uncommon to get females instead of males. Or for random and bizarre symbols such as machine parts, unidentifiable logos, or baby faces to show up if you prompt for a different type of shirt or a person with a smaller or larger stature. Or to get blue shoes when you asked for red.
Try and change the expression on someone's face, and you randomly get close-to-photo representations and an equal number of animes, blobs of nothing, nothing at all (face removed) or some other kind of ridiculous rendering like table lamps or plants.
Good luck trying to get an image sharper. Prompting with "Remove" or "Background" or "Delete" seldom works, but if any of them do, it is purely random, as the prompts are not consistent.
Ask for something to be baby blue, and it will invariably produce a baby somewhere in the picture.
I've had renderings of people with two left hands, elbows positioned the wrong way, and head upside down. I tried to prompt the system that my image has two left hands, to no avail (then again getting hands with the right number of fingers is a real rarity). Not "thumb on other side," not "a set of hands" not "mated hands." I tried prompting that the subject's head was on backwards, again no help. I tried to insert sneakers on a man and been confronted with 250 or so insertions of stilettoes, often times backwards with the heel in front.
I asked for paisley clothes and got solid. I asked for solid clothes and got tiedyed. Hundreds of submits and eventually one comes along that is barely passable, but it is not actually good work that I can eventually use.
Changing the backround almost invariably degrades the frontground image, and not infrequently adds stuff that doesn't belong.
The videos that Jesus has been dropping invariably show first result perfection with some change of the scenery, clothes, etc. This would seem to be so vanishingly rare to the point of not being true. As a for instance, he changed the pants on a subject from chinos to jeans, and it worked perfectly, with the subject wearing what seemed to be 5 minute old pants with perfect texture. I tried it hundreds of times with, candidly, not brightly lit ultrasharp high def photos, but I never got anything to even remotely look like real dungarees. Usually I got a mosh of just blue that seemed to have been spray painted onto the subject's body with no hint of denim channels.
Am I doing something wrong? Am I expecting too much based on the hype from the company? Can any part of the process really come under control so using the app can be productive ie profitable?
AdChoices