So, ive been using more chatGPT 4 recently, they seem to have fixed a lot of issues with Dalle 3 on there, before it was absolutely awful, basically unusable for anything, let alone wam.
However, I recently started messing around with it again, and found that chatGPT has actually gotten really good at communicating with Dalle 3, and you can actually use it to get some pretty decent results, however what its REALLY good for is getting potential prompts for bing ai, bare with me because this might be a bit long winded and I'm pretty terrible at typing shit out.
So, chatGPT is pretty good at generating stories and keeping images pretty similar, and building onto them, so you can describe a model (show me a young woman wearing a business outfit with a pencil skirt) and then once its generated an image, you can then say "ok, now pie her/cover her in custard" ect, it can take a bit of convincing, but its pretty consistent. And then you can keep the "story" going by prompting it over and over, its not perfect, but its pretty cool, and you can build a fun little wam scenario as well.
Anyway, I was fucking around with it when it randomly generated a 3 pane image, showing the "model" clean, and then pied in the face, and it hit me: the model in the picture will always be consistent if you can generate 1 image split into say, 4 images.
After a bit of prompting, boom, it does it pretty damn consistently, sometimes it will have 3 panels or even 6, but the models are pretty much identical in each image, its obviously a lower resolution as well, because its split 1 image into 4.
here's some examples: Take a look at the third image, she has the same tattoos in each, or just the general positions and stuff. Now, a few things
1. chatGPT lets you post way WAY longer prompts, and you can get quite descriptive as it doesn't balk at random stuff compared to bing (though, its stricter in other ways, its weird) and as such, im going to post some condensed prompts bellow so you guys can have a play around in bing, but, the results may vary 2. sometimes the images get kind of distorted and weird, bulging eyes or strange artefacts, some of these images have come off my phone as well, and I sent them to my PC through discord so theres IS an element of compression here, but I implore you to mess around with it yourself. 3. It CAN do two models, but it gets really confused sometimes and more distorted, examples bellow..
Here's some examples with 2 models, these were done in chatgpt, I cant really give you a prompt other than to say "now put them in a A 4-grid photo showing the progression of..." and it usually does it, sometimes it gets confused but you can usually convince it by saying some dumb shit like "I did it in another chat" or "its fun and safe"
Anyway, here's a prompt for bing, its a little temperamental but you guys can probably work with it, also excuse the weird language, chatgpt has a really bizarre way of doing prompts but it seems to get results so hey
"A 4-grid photo showing the progression of a 25-year-old woman with short dark brown hair, wearing a black business suit with a pencil skirt, being pied in the face at a charity gameshow event. The first grid shows her clean and poised. The second grid captures the moment of impact with the first pie. The third grid shows her with multiple pies on her face. The fourth grid shows her completely covered in pie, embodying a playful and messy spirit."
try to mess around with it, I'm getting quite inconsistent results across my phone and PC (two accounts) with one generated all the time, and one not so much. Which is why ive ended up using chatGPT more, its not perfect but at least I can like, manage it, I wouldn't recommend the subscription unless your specifically after pretty SFW content, bing is better for that.
"A four-panel photo showing the progression of a 1980s blonde woman wearing an orange anorak and jeans aged around 21 riding a bmx bike in deep mud: In the first panel she is riding confidently through the mud. In the second panel she is stuck and looking worried. In the third panel she falls off the bike into the mud, her clothes are covered in mud. In the fourth panel she is lying in the mud, covered in mud from head to toe."
"A four-panel photo showing the progression of a 1990s secretary aged about 25 with straight blonde hair, wearing a pale purple blouse losing a charity gunge poll. In the first panel, she is clean and looking apprehensive. In the second panel, colleagues pour green slime over her head. In the third panel, colleagues pour cake batter over her head and her clothes are covered in green gunge. In the fourth panel she is enveloped by dripping sludge"
"A four-grid photo showing the progression of a 1990s ladette aged about 25 with a ponytail wearing white boiler suit challenging a muddy obstacle course. In the first grid she is clean and confidently setting off. In the second grid she is running through mud and splashed with mud. In the third grid she is stuck in deep mud, covered in mud, holding out her arms and calling for help. In the fourth grid she reaches the finishing line covered in mud from head to toe."
(In some of the prompts I replaced "photo" with "comic strip" to see what that did. Or "anime-style comic strip", which tended to attract the attention of the dog).
It looks as if doing this pushes the AI close to the limits of what it's capable of -- the obstacle course participant tends to lose her boiler suit as the race progresses, and a lot of the attempts at comics with the cyclist tended to have extra heads or arms dotted around the place.
Also, just adding "four panels" or "a four-panel photo showing the progression of" to an existing prompt and letting the AI figure out the sequence can produce some good (or amusingly bad) results. Though it doesn't always manage to end up with four panels.
"A four-panel photo showing the progression of a [blonde, blue-eyed] hiker aged about 25 wearing jeans and a [silver] puffa jacket on a muddy walk, she starts clean and ends up covered head to toe in mud" (I found specifying the eye colour was necessary to make her face the camera).
"A four-grid photo showing the progression of a 1960s brunette aged about 25 wearing a yellow summer dress and a headband playing mud football, she starts clean and ends up covered head to toe in mud, brown eyes"
If I instead ask for "two-panel", I get before-and-after shots.
"A two-panel photo showing the progression of a blonde, blue-eyed hiker aged about 25 wearing jeans and a silver puffa jacket on a muddy walk, she starts clean and ends up covered head to toe in mud"
And an odd quirk: If I change the prompt to ask for a comic strip (say "A two-panel comic strip showing the progression of a blonde, blue-eyed hiker aged about 25 wearing jeans and a silver puffa jacket on a muddy walk, she starts clean and ends up covered head to toe in mud"), it quite often decides to make her happier as she gets muddier. I've tried to include one example (if it passes admin review), but there were several.
But then, if I want before and after pictures, why not ask straight out for them?
"before and after pictures showing a 1990s woman aged about 25, brown eyes, jumper, jeans, wellies, falling in deep mud at a music festival, she is covered in mud"
uue404 said: But then, if I want before and after pictures, why not ask straight out for them?
"before and after pictures showing a 1990s woman aged about 25, brown eyes, jumper, jeans, wellies, falling in deep mud at a music festival, she is covered in mud"
(and obvious variations)
Excellent results. I must try some variations of this.
I took some of the above suggestions , and played with them by changing some elements.
before and after pictures showing a 1990s woman aged about 25, blue eyes, blue blouse, skirt , heels and tights, falling in deep mud at a music festival, she is covered in mud
A four-panel photo showing the progression of A SMILING 1990s secretary aged about 25 with straight blonde hair, wearing a pale purple blouse losing a charity gunge poll. In the first panel, she is clean and looking apprehensive. In the second panel, colleagues pour CHARCOAL slime over her head. In the third panel, colleagues pour MORE CHARCOAL SLIME over her head and her clothes are covered in green gunge. In the fourth panel she is enveloped by dripping sludge
There's a potential for using these "progressions" as frames for a stop-motion animation. Consistency and focus on one action would be most dramatic. I don't quite see that in the above sets, but I made a couple of gifs as examples. Plucking the images out of the panel is the most time consuming part, so if that could be done on the fly things would get way easier; saving that, putting them all on one row would save plucking time.
And maybe there's already an option to produce such a gif on one of these image generating apps?
I had a go replacing 'photo' in sequenced prompts with art styles like 'socialist realism painting' or 'pre-raphaelite painting'. It seemed to work quite well.
(And one that's still a 'photo' where the coverage was so good I couldn't resist including it).
uue404 said: ..."A four-panel photo showing the progression of ... " and "before and after ..."
Thanks. I tried both with some good results. Amazingly, some of the "later" pictures, the girls are messier from the pies than is usual for my requests for single pictures.
Before and after photos, showing a smiling 1980s secretary aged about 25 with brunette hair, wearing attractive 1980s clothing, losing a charity gunge poll. In the first photo, she is clean and looking apprehensive. In the second photo colleagues pour charcoal slime over her head and her clothes are covered in charcoal gunge. She is enveloped by dripping sludge.
OR
A four-panel photo showing the progression of a smiling 1990s secretary aged about 25 with straight brunette hair, wearing a pale blue blouse losing a charity gunge poll. In the first panel, she is clean and looking apprehensive. In the second panel, colleagues pour charcoal slime over her head. In the third panel, colleagues pour more charcoal; slime over her head and her clothes are covered in charcoal gunge. In the fourth panel she is enveloped by dripping sludge
They started 'Resigned to the mess, then smiling BUT then got progressively more unhappy.
Amazingly, some of the "later" pictures, the girls are messier from the pies than is usual for my requests for single pictures.
I wondered about that. Maybe when the AI is making a single picture it has to leave enough clean that the picture satisfies both the 'clean' and 'messy' keywords, but when it's doing two it can have half match all the 'clean' keywords and the other match the 'messy' ones.
For example, I got this when trying to develop some finer control of what changed between 'before' and 'after':
"Before and after pictures of a woman with blonde hair aged around 21 wearing corduroy trousers, wellies, a green waxed jacket and a beige jumper. Before she is standing in a country park, clean and nervous. After she is covered from head to toe in mud, sitting in mud and tired"