I don't know what it is but all the generators I'm using tend to have the BACK of the pie hitting the woman's face (but splattering). Since I wasn't doing anything else today I worked on changing that. Best I've come up with so far is "pulls a now-empty pie tin off her face after the cream pie has splattered all over her face and chest."

Some success.

The last picture not so much but I liked her "attitude".

I've mentioned before this is one of the major challenges with LLM image generators, and I think it's difficulty lies in the program attempting to show a recognisable face whilst also being asked to partly obscure it with cream/mess. Given that the model is also replicating the pool of all available existing images and the vast majority are clean faces, you can start to appreciate why making a pied face is hard.

The reason it's difficult to generate is pretty straight forward. The training material doesn't include enough material to recognise what a pie in the face should look like. There is no getting around this fact. Specifically trained models for flux, etc show no issue with the subject.
The fact is, there is a compromise using public models

So is there a way to get this to work without spending a lot of money on Generator subscriptions and doing a lot of programming? Are there free or inexpensive generators that allow you to upload an existing picture to show what you want?

It's not clear from your message if you mean it's difficult to have the pie hit the face from the right angle / orientation or to have the character removing / wiping the pie after being hit.

Getting the orientation right should be fairly consistent (around 80% of the time) if you use Sora or ChatGPT. It's easier if you have a character pushing it but it works either way.

With Bing / Designer it's a lot more random as those tools struggle with complex positioning. It's possible but requires more attempts and luck.


Getting the character to remove the pie adds a lot of complexity because the generator cannot imagine an object being at two places at the same time. It will try to remove the pie from the face and the result will be a lot less consistent or pleasant. You can trick the system by describing two different pies.


I've attached a few examples, explain what you want to do and I'll give more detailed instructions.

Both. ImageFX in particular has the girl using a cloth to wipe off the pie, even when I've specified using her hands. I really like the victim wiping the pie off her face and cleavage.

I'm using Krea on a low-level subscription (but the images created have been "over-exposed" lately); ImageFX, Ideogram (low-level sub) and Gemini.


I've all but given up on Bing; it often just spins and nothing is generated. I haven't tried Sora or ChatGPT yet.

I really like your 2nd and 3rd pictures here.

My favorite "pie deliveries" are the flying pie out of nowhere and pushed/smashed by another woman, preferably one who has already been pied.

Thanks for your help.

No problem, here are the prompts for Sora / ChatGPT. I am not sure they would work on ImageFx until it is updated to use Imagen4.

Flying pie from the side:

Ultra-photorealistic slapstick comedy image, 3/4 angle, cinematic lighting.

A glamorous woman in (outfit) is captured at the precise instant after being hit full in the face with a cream pie.
A large, clean, slightly dented pie tin is pressed tightly against her face, following the contours of her nose, chin, and brow, perfectly hiding her features except for the top of her messy, cream-overwhelmed hair visible above the rim.
She is shown from a dramatic 3/4 angle.

Key Impact & Movement Details:
Her head is visibly recoiling backward from the force of the pie impact--neck arched, hair and ponytail flying back, and shoulders pulled slightly away from the point of contact.
The posture and tilt of her head and body clearly convey a sudden, comical "knock-back" at the moment of impact.
The tin's dented shape closely hugs her facial contours, accentuating the direct hit.

Splatter Direction:
From the exact point of impact, a huge explosion of thick white whipped cream and glossy pie filling bursts outward and backward, forming a spectacular, wide halo of mess behind her head and shoulders.
Cream and filling arc dramatically behind her, with large globs and droplets frozen in midair to capture the slapstick energy.
Heavy ribbons and streaks of cream wrap around her hair, neck, and upper back, with extra splatter flying beyond the edges of her recoiling head.

Additional Details:
Her glossy outfit is streaked and splattered with cream, especially over the chest, shoulders, and arms, with mess trailing backward due to her head's movement.
Her hands and arms may be lifted slightly, further suggesting the surprise and force of the hit.
The background is softly blurred--an elegant setting, perhaps a ballroom or gameshow--with warm, cinematic lighting that catches the flying splatter and glossy cream.
The whole image radiates slapstick energy: a pristine pie tin, a powerful head knock-back, and an unmistakable explosion of cream bursting behind her.



For Sora:
Show a woman just pied in the face from a 3/4 angle, with her head visibly recoiling backward from the force of the hit--neck arched, hair flying back. The clean, shiny, dented tin is pressed to her face, following her features. All the cream and filling explode outward and backward from behind the tin, forming a wide halo of flying mess behind her head and shoulders. Her glossy outfit and streaked hair are drenched in cream, and the whole scene is dynamic, comic, and ultra-photorealistic.

If you copy this version it works pretty much all the time, although you won't always get the 'clean back' effect all the time. It also works with different angles and materials.


For the second one with direct shove, there are a million different versions but this one is kinda consistent:

A high-resolution, ultra-photorealistic slapstick comedy image, shot from a dynamic over-the-shoulder third-person perspective--angled just behind the woman receiving the pie, to emphasize motion and dominance.

One glamorous guest is energetically delivering a giant classic cream pie--slightly larger than the recipient's head--directly into another woman's face. The moment is captured at the exact instant of gooey impact. The pie is rendered as two distinct elements:
A shiny silver pie tin, viewed from behind as it crumples under the pressure of her palm.
A mass of ultra-thick, heavy whipped cream, enriched with marshmallow fluff for maximum cling. The cream is **dense, glossy, and sticky--not airy or foamy--**and erupts in heavy folds and splashes as it hits.

The pie-throwing maid's dominant arm is fully extended, her open gloved palm pressing directly into the center of the tin, ensuring a straight, centered hit. The pie is delivered with authority and confidence, producing a loud squelchy SPLLURSSSH sound.

The woman receiving the pie is on the floor dramatically recoiling--shoulders jolted backward, her body reacting to the overwhelming pie impact. The thick cream engulfs her face in a smooth, flowing curtain, while a satisfying secondary blast splashes across her neckline and glossy bodice, with globs visibly flying outward. Her styled hair is already soaked, sticking into the mess, and her expression is frozen, utterly overwhelmed.

In the background, elegantly dressed guests are mid-laugh, ducking or cheering, with airborne pies and flying splatter everywhere. The lighting is bright and theatrical, catching every drip, gloss, and detail in the chaos.

This one is also fairly model-independent and should work with most character descriptions or models.

Let me know if you need explanations on some of those. You can check my previous post if you need info on using models / presets.

Wow. Thank you for that detailed prompt and all the work that went into it. It seems to work!

I revert to what kortanklein said. You really need to understand the nuances of how each model works. Sora/GPT uses a different method to generating images that usual diffusion models which makes it better interpreting prompts but you still need to learn the nuances of the prompting.

Most other models are diffusion based and will have difficulty to certain degrees. I'd ditch Krea if I was you. Historically, their models were just based on SDXL, FLUX etc with a few tweaks but still don't compensate for messy. Their new model is self trained but still has similar limitations to the image set and prompt understanding.

Flux and Hidream etc work when trained as we're able to train on a much more diverse image set to tailor the models and prompt understanding. The output will be generally better quality but limited in range to say GPT-4o/Sora.

Imagen 4, I believe uses a similar generation method to Sora but I'll be honest, my testing with Imagen4 hasn't resulted in anything as good as Sora. Simply put Sora/GPT4o is the best you are gonig to get from public models just now.

Thanks! Off to try out those new prompts and maybe new generators.

Some early results. Of the three generators I'm presently using, Gemini did the best with these prompts. ImageFX and Ideogram also improved. However, I dropped the phrase "ollowing the contours of her nose, chin, and brow" because it gave me the results in the next-to-last picture, with her nose and lips poking through the aluminum pie tin! The 4th or 5th picture is the same prompt but without that phrase.

Thanks, I've had a little more success after reading your prompts, although been better at mess dumped on the head rather than directly in the face.