I've been frustrated in attempts to get dunk tanks in images. The LLM generally ignores the prompt and puts the person facing the wrong way, or the seat floating over the middle of the tank, or the tank is too small.
Gloopsuit said: I've been frustrated in attempts to get dunk tanks in images. The LLM generally ignores the prompt and puts the person facing the wrong way, or the seat floating over the middle of the tank, or the tank is too small.
You need to think laterally about what you're trying to achieve and what context does the model understand. It is highly unlikely that it understands what a gungetank is. It also doesn't understand how a person would sit in it. You need to to be as specific as possible within the confines of the prompt. If you are using GPT4o or Sora, you can sketch and image in pencil and upload it with specific instructions. It is smart enough to understand your sketch and notes.
I'm getting consistent dunk-tank image results in Whisk, specifying an elevated over the shoulder perspective. It's the video prompts that are infuriating me, the entire seat assembly constantly falls with the contestants attached to it, regardless of how I've described it. It seems like the only thing Kling can generate successfully is a credit card deduction.
Edit - It did just manage to do this, but it cost thousands of credits to get there.
Gloopsuit said: I've been frustrated in attempts to get dunk tanks in images. The LLM generally ignores the prompt and puts the person facing the wrong way, or the seat floating over the middle of the tank, or the tank is too small.
You need to think laterally about what you're trying to achieve and what context does the model understand. It is highly unlikely that it understands what a gungetank is. It also doesn't understand how a person would sit in it. You need to to be as specific as possible within the confines of the prompt. If you are using GPT4o or Sora, you can sketch and image in pencil and upload it with specific instructions. It is smart enough to understand your sketch and notes.
Oh, I have been. The prompt has gone into detail such as 'a plastic seat attached to the inside rim of the vat' and 'she is sitting above the gunge with legs over the edge of the seat.' Didn't know you could use a sketch as a basis, will try that.
Generated these with Veo3 via Gemini. The woman takes/slips off one of her shoes and steps on a Cool Whip pie with her bare foot, and shows her creamed sole to the camera.
Working on cleaning up the prompts so that limb movement is consistent.
Another post of bikini babes getting pied due to lost bets. I have so many kling videos now I dunno what to do with them, and will try to post more if people are interested. Here's the video on youtube below which is higher quality. Enjoy y'all!
I think this magazine might have caught my attention. But which cover to choose?
(These are ImageFX portrait images, so the previews are only showing the top half, and the mud's mostly at the bottom)
I made these by starting from a genuine 1980s fashion photo (not included because that would be a picture of a real person), then using imagetoprompt.com to get a detailed description of the model's makeup and hair. That became the basis for the ImageFX prompts.
So, WAN 2.2 is pretty good at WAM! Went for a "bikini-clad clowns at the circus pieing themselves in the face while wearing Pie Me party hats" theme this time around. Animated gifs (give them some time to load) of the pieings, plus I posted the higher-quality/resolution clip to youtube - https://www.youtube.com/watch?v=m5JJvtyU2Y4 - I hope y'all dig these!
For those that want to try this themselves on WAN 2.2 (free and open source), here's my (probably overly verbose) prompt:
A woman holding a plate full of whipped cream in her hand. She presses the plate full of whipped cream against her face, splatting the plate full of liquid gooey lumpy whipped cream all over her forehead and face, sending whipped cream flying behind her, covering her face, hair and hat in whipped cream. This covers her entire face and drips onto her breasts in a thick gooey layer of wet dripping whipped cream as it drips from her face to her chest. In the end, the woman lowers the empty plate to show her cream covered face, with her whole face (hair brow, forehead, cheeks, chin, eyes) and breasts all covered in white gooey fluffy lumpy whipped cream as she opens her mouth in shock then sticks her cream-covered tongue. The scene ends with the woman's entire face covered in a thick layer of whipped cream and her breasts covered in the whipped cream and the plate empty. Make the whipped cream drip and look really bumpy and gooey, leaving trails of white cream, and look wet and gooey.
Not worthy of a thread yet but I think I am getting really close to replicating the vibe of specific real life events that feature people receiving large cakes in the face.
It's a huge pain in the ass since I'm not using any of the real life event images as model, this is 100% prompt-based.
One of the earliest WaM stories I read had a baker's assistant getting pied 13 times, including 1 to each foot. It was one of the first instances of foot-pieings I read or saw images of (rather than just stepping in a pie.)
Using the "add images" function of Whisk that I discussed in an earlier separate post, I took a wetlook AI picture and added some pie.
After Whisk describes the image, I add something like "A cream pie has hit her face and splattered on her shoulders and chest." or "A cream pie has hit her face and she is wiping the remains of that cream pie out of her cleavage." to the lower mustard-colored window. It doesn't always work; sometimes the original picture is too sexy or looks too much like a celebrity or Whisk is simply out getting a smoke. Sometimes I change "cleavage" to "chest" or "décolletage."
I find Whisk "pies" the pictures better than in its main "describe your ideas" window. The results are also much closer to the original image than from the main window.
I often see a picture of a beautiful woman, dry or wet, real or AI, and think "I wish someone would hit her with a pie." Now I can.
I now have access to Sora 2 so I've started some video tests before running out of prompts for the day.
It is really impressive, especially the sound part which is stunningly natural in every language. I need to figure out how to extract a video without "posting" it but you can see screen cap examples of self-pieing being done in basically half an hour.
It's night and day compared to Sora 2. Sadly you can't do I2V with photorealistic images (even AI-generated).
I tested SORA2 a bit more using my 100 daily videos, there is definitely some potential once you understand what the tool does well and what it does not.
The biggest limitation is the lack of I2V with photorealistic images, really kills your prompt budget since you have to specify everything in the main scene.
Here are a few more proof-of-concept videos: they are kinda bad and full of artifacts and cuts but can give an idea of the tool.
kortanklein said: I tested SORA2 a bit more using my 100 daily videos, there is definitely some potential once you understand what the tool does well and what it does not.
The biggest limitation is the lack of I2V with photorealistic images, really kills your prompt budget since you have to specify everything in the main scene.
Here are a few more proof-of-concept videos: they are kinda bad and full of artifacts and cuts but can give an idea of the tool.
If anyone wants I still have 2 invites left btw.
Thanks for one of the invites - I have a lot less success getting around the guardrails though!
I wanted to try Sora2 to see if it could get the pie throwing physics correct when it comes to pushing the pie into a woman's bare feet. Not there yet with the "perfect pie-dicure", but the "model's" reaction was a lot more lively and cuter than what I got with Veo3.
kortanklein said: I tested SORA2 a bit more using my 100 daily videos, there is definitely some potential once you understand what the tool does well and what it does not.
The biggest limitation is the lack of I2V with photorealistic images, really kills your prompt budget since you have to specify everything in the main scene.
Here are a few more proof-of-concept videos: they are kinda bad and full of artifacts and cuts but can give an idea of the tool.
If anyone wants I still have 2 invites left btw.
Thanks for one of the invites - I have a lot less success getting around the guardrails though!
I wanted to try Sora2 to see if it could get the pie throwing physics correct when it comes to pushing the pie into a woman's bare feet. Not there yet with the "perfect pie-dicure", but the "model's" reaction was a lot more lively and cuter than what I got with Veo3.