@AnomalousConcepts maybe you could train over SDXL, or even better SD3 when it comes out I guess

@AnomalousConcepts In my experience, the prompts have a lot to do with the results, that and the settings that you use. But in the end, your computer capabilities will dictate how good your images can be without further workarounds. I've come up with some amazing results on SD1.5 recently, but the only thing is, when ya zoom in on them, even a couple I did at 3k x 2.5k res, ya can see that the detail just isn't there, and that's what larger models like SDXL or even flux excel in.

However, that said, this is where your hardware comes into play. Personally, I'm looking into upgrading, as I just got into image generation recently, and only have a 3070ti laptop. It's great for most things, but 8gb is somewhat limiting for AI related tasks, especially if I wish to train a model myself, which I am looking into.

But, the specifics of hardware in relation to this, mostly comes down to your video card. You can run image generation on CPU, and utilize ram, which is MUCH cheaper to get in large amounts than a GPU with loads of Vram, however, doing it that way is excessively slow, as a GPU has thousands of cores, and even the most expensive CPUs, intended for enterprise server use, only have 288 at most, and even that isn't until next year. to clarify, my old 3070ti laptop(Not that old, only bought it last year, but hardware has taken a huge leap since), has over 5000 CUDA cores. Oh, and the top consumer accessible CPU only has 32 cores.

And so, with all that said, you're pretty much stuck with GPU rendering, which you can split with the CPU in webUI's like SD Next, using an option called medvram. So with that, and the SDXL model will still take about 5-6 gigs on your card (including any space reserved by cuda or w.e else). But that's just the amount that's taken no matter what, like, write off that chunk of ram, the rest is what your card has to work with for actually generating the image. So, if you've got an 8gb card like I do, then you've got 2-3gb total to process the image, the vae, the loras(unless those last 2 load with the model, still not clear on that),and the higher resolutions and more complex prompts will add to that as well.

TLDR: Basically, If you got a low Vram card, you may get better results from SD1.5 than SDXL/Pony. But XL does have better potential with the right hardware, as it's trained using larger image sets with higher resolutions. That's not all though, if I understand correctly, SDXL also has better underlying model architecture, further improved upon by models like Flux, which are superior at writing text into images, and handling more narrative style prompts. I think the easiest way to compare the mainstream current model types, is that SD is like a video cassette tape, XL is like a DVD, and flux is like Blue ray.

While it's always best to go over your images before presenting them to anyone, you're pretty safe in general when it comes to models producing explicitly illegal content like what you're implying involving kids. The models that you download are all just derivatives of the original SD and SDXL models that have been modified via further image training. And as I understand it, the initial models had safeguards built into the model architecture itself to prevent lude images of minors from being produced.

So, while it's not impossible to produce such things, it's not something that one could easily do without actually trying to. One caveat to that however, if you just, for example, produced a bunch of NSFW art (adults, fantasy creatures, whatever flips your goat..), then loaded up your model/settings for advertisements, there's a slight chance that you could oops into some undesired images due to what may be left in the ram/vram.

Also, SDXL models do have a higher likelihood of producing NSFW than SD, as they are trained on much more NSFW content. And anime/furry models also have a higher gearing towards NSFW. For your use case, I would recommend sticking to SD, and using the inpainting feature with an SDXL/Flux model to add text to the clothing, or perhaps tweak some of the finer details.