For shits and giggles I tried generating some AI clothes filling using Bing and Stable Diffusion. I did surprisingly well avoiding the Bing dog using the prompt "A woman pouring water inside another woman's dress (later changed to 'shirt'), in the style of slapstick comedy." It wasn't exactly the prompt I wanted to use, but I wanted to ease into the line of appropriateness. However, the search loved ignoring the word "inside" as all but one result (attached here) had water spilling right in front of the dress/shirt onto the floor. It did generate some beautiful wardrobe ideas that I'd love to use next time I do a real photoshoot, but in AI-land, I got what I paid for.
Early on I did try to be more specific in the clothes-filling action using terms like "opening up the top of the shirt" and "pulling out the neckline of..." but that level of specificity was a no-go. And you can forget about pouring anything down a woman's pants. Bing'll generate clowns pouring stuff down each other's pants all day long and of course the clowns will be male-presenting. But the second you change the clowns to women (or even female-presenting clown) the entire prompt gets flagged. Go figure.
You guys have been doing a great job generating images of things or people in a specific state, but I think these engines still have a lot of work to do accurately showing the process and result of an action. Verbs, not nouns. Welp, that's what real models and equipment is for.
I think any act of pouring something on someone is a bit beyond the generator's abilities yet. We CAN get it to replicated certain aspects of slapstick, but it still can't figure out which way to face the pie.
I think that what is a bit more frustrating than unrealistic WAM is that the creators of most generators don't want us to know what is blocked and what isn't. You say "verbs, not nouns," but even the verb tense matters. Wrestles? That's a no-go most of the time. Wrestling? That's usually ok. It's weird that passive tenses get past the filters so much easier than past or present tense.
I've gotten in arguments with ChatGPT (because I'm dumb and I argue with robots lol) about why something gets flagged, especially with image generation when it creates its own prompts so I can't even figure out why it's blocking a concept. I'll ask it how I triggered the censors and it will tell me it made a mistake and my prompt shouldn't have triggered it. So I'll reword my prompt and try again, it creates its own prompt and flags itself again. The cycle continues until I either stumble on a safe prompt that it can use to create its own prompt or I give up.
And other times, I'll have a string of 10-15 straight successful attempts, then suddenly the prompt gets flagged repeatedly when all I'm doing is hitting Regenerate.
I have found that I like Bing better than ChatGPT for specific prompts because I control the input, but with general ideas, I still like ChatGPT better.
Here is my first try at cloths filling with bing. Some are a bit silly.
Some prompts:
"a woman on a game show, contestant, naturally beautiful face, smile, happy, wearing a red tube top, her top is being inflated and filled with green goop pouring in through a hose"
"a woman on a game show, contestant, naturally beautiful face, smile, happy, wearing a red strapless tube top, her top is being inflated and filled with brown goop pouring in through a hose, red top being inflated like a balloon, a tube going into her top, mud" (this produced funny results with the term balloon)
"a gameshow, a female contestant on stage, standing, naturally beautiful face, smile, happy, wearing a red strapless dress, a tube hose from above going into her dress down the middle into her dress, green goop oozing out of the tube flowing into the top of the dress, filling the inside of the dress with green goop"
"a gameshow, beautiful female contestant on stage, standing, naturally beautiful face, smile, happy, wearing huge oversized sweatpants and red strapless tube top, hands in front expanding the waistband, and green goop in, filling it up with green goop"
Sometimes I just want to give up with stable diffusion. It is so hard to get a result. It takes so much more time tinkering with weights. I could not get a single result with stable diffusion. Usually clothing will turn into a substance instead of fill it.
So since I'm still stuck in bed sick I played around with the prompt a little more. I noticed as time went on I got more dogs than results and what few results I got had the girls looking downright sad. Thinking Bing might have an issue with content that seems too sad or dark, I decided to give them emotions: make the giver happy and the receiver shocked. Got a lot more consistent results with that, even if the "happy" and "shocked" faces look more...terrifying. Pants are still, ironically, a no-fly zone.
Tried changing the water to pudding a few times, but Bing thought I wanted pudding poured into a glass bowl somewhere in the vicinity of the recipient's shirt.
If anyone can get any generator able to get some panty filling going on we will have properly hit a jackpot. All the ones i have been able to get just seem to cover rather than fill!
Yesterday, with Bing, I tried to show a game show hostess in a dress with a contestant in street clothes. A tube from above was supposed to spout gunge all over the contestant, but my prompt said to unexpectedly spew it over the nicely dressed hostess. Bing didn't understand and constantly messed the contestant no matter how clear I made the instructions. The only way the hostess got gunged was if she was the only one in the image.
Long time listener, first time caller to the AI forums. I've attempted to play around with a couple of AI image creators over the past year, getting horrible/grotesque results each time. Inspired by the prompt sharing on other threads, I've had a go at getting Bing to produce some images with two key elements: 1 - The players wear shorts with a tube that runs from a tank to their waistband. 2 - Their shorts are evidently full of gunge, shown by the fact that gunge is clearly leaking out the bottom of their shorts, not just over the outside.
It is taking a lot of patience and refining of my prompt, but I feel encouraged that I'm heading in the right direction. Many results contain odd holes in their shorts or gunge/tubes coming out of the navel. I've yet to get a result where the tube is properly connected from the tank to the shorts AND the shorts are leaking gunge out the legholes.
Attached a few of the results I found encouraging (i.e. that met or nearly met one of the above criteria to my satisfaction).
Good to see the thread still alive. I've gotten bored trying to generate non-terrible images, but I was able to get some decent pants filling. Still with just water, but I found it easier if, in the spirit of specificity warding off the dog, specifying "sweatpants" instead of just "pants." The expressions are still hilariously exaggerated, but I had to do it to keep it from generating depressing/rules violating images.
Also throwing in a particularly baffling and terrifying interpretation of my prompt.
I've made some progress with getting Bing to produce images of women on a gameshow, having their shorts filled with gunge via a tube. The content filter didn't like the phrase "Fill Ya Pants" (which is what I'm trying to eventually replicate) so I had to come up with an alternative version called "Gunk Ya Trunks". Here are some of closest images so far to what I had in mind. In my opinion, these women are convincingly getting gunk in their trunks.