>>19961I didn't use a base image for this one, no. But I certainly hope this isn't a drawing that already exists! If you do happen to find it, let me know...
>>20135Glad you like it.
>>20159I can say the idea has occurred to me, but I have to admit I'm not very interested in it. What you're proposing sounds like a limited toolset art challenge, like "Create a painting using only four colors and the clone stamp tool." I'm sure someone would be able to do something cool with a challenge like that, but it's not for me.
This is something I've talked about before regarding the way we approach StableDiffusion and other image 'generators': nearly everyone thinks of them like magic image machines where you type in the right words in just the right way and then they give you exactly what you want. I honestly believe that's not going to happen anytime soon, and it's definitely not how they work right now.
You've touched on the primary reason why, though: there isn't enough data for something like that in most models. I'm not an AI expert, but my understanding of how this all works is it's effectively auto-complete for collage work. When you type something like "beautiful woman, large breasts, bikini, pool" SD trawls through a context database to find images (or rather the concept of images, but we're getting into the weeds here) that share most of those tags. after generating noise, it tries see patterns in the noise and 'draw' lines and colors that have similar shapes and values to the images it's referencing that connect the points it 'thinks' it 'sees'. It's not literally doing collage work (cutting parts of images out and gluing them together), it's a bit more like tracing. But at the end of the day, SD can only 'trace' things it's 'seen' before. That's why if you use the example prompt above, you'll get a pretty nice looking image even at a low resolution: the core SD model has seen billions of images of conventionally attractive women in bikinis with 'large' breasts, so it's really good at spitting similar images out. If you type "A 1000 pound woman squeezes through a doorway, gigantic belly, doorway, stuck" it's going to loose it's mind trying to make sense of those words.
That's also why there are so many specific checkpoints that have been created for use with SD. People are basically creating new context sets for SD to utilize. To further stretch my collage metaphor, say you wanted to create the image you're describing by cutting out pictures from magazines and gluing them together. If I hand you a stack of Maxims and Playboys, you're going to have a pretty hard time making it work. But if I gave you the full print run of BUF and other BBW focused magazines, you'd be much closer to your goal.
This is why I utilize the img2img functionality of SD. If I can take an image I already have (whether it be a real photo or a rough image generated with SD) and do some fairly simple edits to it, I can feed that back into SD and tell it to keep that shape in mind. The following generations already have a base image to reference, and it can recognize elements of the image to work from.
So, until we live in a world where the average woman weighs 800 pounds and everyone gets turned on by seeing them struggle to fit through a doorway, we're just not going to have enough resources to pull from for SD to be able to create a similar image accurately from only a prompt. Even when you include art (while not a massive dataset, still enough to get the general idea of a massively fat woman across) what you're looking for is pretty niche, and I would (generously) guess there's only a couple hundred drawn images to draw from, AND that's assuming they've all been compiled into a LoRA or a checkpoint somewhere.
So, the image I've attached here is an example of what I was able to do in the past. Using (a now old version of) Symix, the creator of that LoRA was able to tag some images with "stuck_in_doorway" which I call in the prompt:
>a fatblob woman is stuck in a doorway, stuck, stuck_in_doorway, fatblob, gigantic belly, huge breasts, fat face, two tone lighting, dimly lit, low key, <lora:epiNoiseoffset_v2-pynoise:2> <lora:symix-preview-4-9:1>Out of this grid of four images, only two of them have the 'pressed against the doorframe' look, but none of them have that 'squish' you're looking for. the next image is the final result after editing the third image in photoshop and reprocessing it multiple times. I had to edit the belly to make it larger, and curve the sides to give the impression of her fat spilling out 'through' the doorframe.
That's also how I would approach your challenge. Rather than try to get the prompt to give me exactly what I want, I get something that's in the vicinity of the image I want to make, then edit it and feed it back into SD. Repeat this process until you have an image you're happy with.
So, I hope that helps somehow. I've included a link to the Symix LoRA that includes a "stuck in doorway" tag, but understand that it won't just magically work. I'm just handing you a magazine that has a few images of fat women stuck in doorways, it's up to you to make it work.
And finally, what checkpoint/LoRAs did you use to generate the images you've shared? I don't recognize the style/shapes, and I'm curious what you did to get them.
To everyone else, a new random image for you.
