What is the most detailed AI xxx character creator?

[reddog2711]

I'm sorry if this is already covered somewhere and I missed it, but is there a VERY detailed AI character creator? Let me explain what I mean.......

A few nights ago, I 'may have' been a little high, and got saying to nude pics: her tits are too small. Hers are too big. Terrible ass. ect.

I had the idea that I'd like to build, what I think is the perfect nude woman from scratch, but far more detailed than just the basics.
I'm hoping that there's one that someone knows of, even if its paid only, where you can choose every little detail, down to nipple size and placement, areola size, same with every detail about the labia, and everything else.

Does something like this exist (yet)?
Thanks

[paulai]

Its possible to create the perfect woman of your desire, but what you're asking for would take about 2 hours for someone as trained as me just to create a single picture. Its almost like painting/drawing her yourself. I normally generate the base woman, then if there is something I don't like, I go over and mask inpaint the face, nipples, vagina, feet, etc like 50x if I want the perfect picture. If you want it perfect then its going to take many hours just for a single picture.

You talk as if you have very little to zero experience using AI tools, so for you it would likely take a very long time, months if not years to get the skills to generate the woman you desire. I started in 2022 and its taken me years to finally get good at generating AI women. To be honest and its going to hurt a lot of peoples feelings, but a lot of the AI 'art' here is low quality and low effort. If you want to see the professionals there are plenty of other AI websites to visit. Some of these pro AI dudes make my stuff look like child's play literally.

A lot of people and real artists trash Ai art generators like me for our stuff, but it really takes a lot of time crafting a good picture, its almost like drawing/painting yourself, there is real skill and craft that goes into it. Sure, if you just want to generate a picture of a womans upper torso with all her limbs cropped out then thats super easy, anyone can do that easily and get good results because its almost impossible for AI to screw that up. The mastery is in full body shots, its why I particularly focus on it due to its difficulty.

TL;DR- Currently there are no easy tools for what you want. AI, particularly NSFW/Nude/Porn AI is still very underdeveloped due to its taboo nature. The mainstream AI industry outright refuses to throw money at training a NSFW AI model because they dont want to be cancelled.

[Neobutra2]

Hi,

if you want the most control, I recommend you do the AI generations by yourself locally on your own computer. However, the best quality is by far, hands down, offered by cloud computed services. There is no question about it, one single nVidia B100 used in AI cloud computing is in a totally different ballpark than over 30 5090 Ti Super Extravaganza Gämer GPU cards. However, these cloud services are limited, as @paulai said above, by controlling what you are allowed to generate, because nipple is holy in the US of A.

But if you can live with ok'ish quality, Stable Diffusion XL (SDXL) is a tool from the lightweight end and it can be run with a graphics card that has 6-8Gb of VRAM memory. Flux (F1) is your next step up from SDXL, and it requires about 12-16Gb of VRAM (you can run it with 8Gb but generations times are huge for one image, and you usually have to generate a few to get one good one). If you have 16 to 24Gb, you can run HiDream, which is at least on par with Flux, and some consider it even better. These are called architectures, the architecture is your first choice you have to make. I'd choose HiDream or Flux if I had a beefy card, and SDXL only if you have an older card with limited VRAM.

With all these architectures, you can affect the image generation by using a certain Checkpoint. Checkpoints are your main training data sets, and there are checkpoints for anime, there are checkpoints for art and there are checkpoints for photorealistic images and so on, some are more contrasty, some more dreamlike, and you can mash them up to get somewhat new mixed results. Some checkpoints are trained with a huge amount of nude imagery, some a bit less. Some are trained with fantasy images, some with dark and moody images and so on. Checkpoint is your main art style choice. You can download these from Civitai, Seaart and Huggingface, for example.

Then you can affect the image generation with LORAs, which are like really specific scenario data sets, something to aid you generating really nuanced things, like puffy lips, bigger eyes, huge labia, innie pussies, larger areolae, puffy nipples, tuberous tits, enormous Zena-Fulsom-like tits, plump bellies, belly jewelry, albino look, and so on and so on and so on. There are tens of thousands of LORAs and you can train your own if you have a beefy graphics card with a lot of VRAM (by using Kohya-SS or OneTrainer for example). You can download these from Civitai, Seaart and Huggingface, for example.

Now when you have your base model architecture (SDXL, Flux or HiDream) selected, and you have downloaded a bunch of checkpoints and LORAs, you start to seed hunt: This means that you run the image creations with just the necessary settings without any high resolution tools, so you can generate images in bulk and fast. You test different prompts and how they affect the results, you tune the weights for the LORAs and different prompt keywords, and you start to get probably what you want. For every creation you get a seed number, which means it is like an address where you will almost always find the same exact picture again and again IF you keep your prompt and weights and LORAs and checkpoint and dimensions exactly the same. Don't worry, your images you create will carry the prompt data with them as metadata. You can feed an earlier created image back to your AI creation software, and it shows you what the prompt, weights and all other stuff was. You can then grab the data straight to your creation tools and create it again but this time with way higher resolution which is the time consuming part. This is usually called a Hires.Fix, but there are other similar techniques as well.

Now that you got a high resolution "end product image", you can still go further and enhance it the way you like. You didn't like her labia? Her areolae were wrong? You wanted a more slanted and seductive eyes? You can inpaint those parts. Inpainting means you bring the image into the AI software but instead of text-to-image you bring it to image-to-image side, and there you can mask the part you want to create again, and prompt and use LORAs which this time will only affect that masked part of the image, the rest stay unchanged. However, if you want really detailed results in inpainting, you need to cheat a little bit. All AI architectures have their max resolution: SDXL is 1 megapixel, Flux is 2 megapixels and HiDream is said to be 4 megapixels. This means that if you try to inpaint an image and affect to a really small part of the overall image, like nipples or labia, at least in SDXL you will have to cut that part out of the image and bring that smaller section into the inpainting, so the affected part is a larger area compared to the canvas (for example, in an image with dimensions of 3300x1900px, a pussy in a full body framed image is only like 1% of the whole image area or even less, but if you cut out a 800x800px area from around her pussy, it is now way larger compared to that 800x800px canvas than it was compared to that 3300x1900px canvas). When you have rolled a good result, you have to attach the inpainted image part back to the original image. I use Photoshop for this, where I set guides to align the image parts to seamlessly attach them back into right position.

So, instead of having a nice website where you can create a high resolution RPG style character who has 65000 different eye styles to choose, 65000 different nipples and so on, you can basically do it in the inpainting section of an AI software. But it is a time consuming process and randomness plays a role. If you have played RPGs, I'd say rolling an AI creation is like throwing a D100 with two ten-sided dice and hoping you get a result over 70 to succeed, but using LORAs and inpainting and other tools it is the same D100 roll but this time you have a +25 Ring of Creation and +20 Cloak of Latent Spaces to give you better odds

On top of these mentioned here, there are lots of other AI tools you can use to further control the creation process. @paulai for example uses Controlnet a lot and excels using it, it is a way to tell the AI generation process that I want this particular pose, or you can give it a depth map of some other image to work with as a reference and so on, it has like 20 different options. Then there are ways to tell the AI generation process which things you want to certain positions in the image, and many more.

- N

[Neobutra2]

I forgot to start from the absolute beginning, which is the AI creating software itself. If we talk about Stable Diffusion, you have many options which can be found online from many discussion threads, like this one: https://www.reddit.com/r/StableDiffusion/..._use_stable/

AI generation UIs

AI generation is often done so your UI is a web browser page (local address). The UIs can vary a lot though, but the good thing is that there are many to choose from so I'm sure you'll find a good one you like.

From those mentioned in the above thread, ComfyUI is among the most flexible ones, as you can run all of the aforementioned architectures with it, and it seems to constantly be on par with the fast development of the different AI trends. However, it has the greatest learning curve of all AI creating UIs I have ever used, and personally, I find it still so confusing that I have not started using it yet, but that is just me. It is by far one of the most popular UIs used for AI creation, and I strongly recommend it. Instead of tabs and setting-boxes it uses nodes, same things you use in Blender or DaVinci Resolve's color and fusion tabs. I'm not that fluent in them, but I admit I haven't studied those enough. If you put a few hours into it, it should start to make sense

I'm a lazy ass, so I run my creations with ForgeUI, because... it is just a simple .exe installer, and after the install everything is basically ready immediately. Everything just works out of the box. All I had to do was download a "few checkpoints" (well, I have acquaired about 450 gigabytes worth of checkpoints and almost 200 gigabytes of LORAs ) and I have downloaded LORAs when I have needed them or wanted to experiment a little. Forge UI can be downloaded from here: https://github.com/lllyasviel/stable-diffusion-webui-forge (browse a bit down and you find the "Installing Forge" section that will guide you through everything).

Also, watching a few Youtube tutorials or reading some github pages from developers works wonders. There is also a lot of precious threads in Reddit, Civitai and Huggingface which tell more ins and outs.

For ForgeUI, I have some basic settings and a guide for enhancing your creations with Adetailer, Hires.Fix and Inpainting:
https://viperproxy.org/threads/9051499-AI...ost235969319

About Checkpoints in Stable Diffusion

In Stable Diffusion, which is open source, there are two main architectures: Stable Diffusion 1.5 (SD15) and Stable Diffusion XL (SDXL). There are also SD2 and SD3, but they are not popular, and people use Flux or HiDream over SD3. The reason is that Stable Diffusion XL and SD3 both use a massive packaging for its training data, so when you generate small things, like an earring or shoe laces, it is a little mushy because the data it uses to generate them is so packed (a whopping x48 times in SDXL).

The SD15 is the oldest of these mentioned architectures, but it has a massive amount of LORAs made for it, which is the main reason people still use it a lot. Also, SD15 requires so little VRAM to run that people are using it with very old graphics cards. Its resolution is less than one megabyte though, so the generated images are really small, around 512x512 pixels, so when you see really tiny stamp sized AI images, they are most probably created with SD15. They can be upscaled though, so you can make larger images, around 1300-1500px, with SD15 as well quite easily, but it requires some extra effort.

SDXL is divided into three popular "finetunes": XL, Pony and Illustrious. XL is SDXL's basic training data which is great for photorealistic images or art imitations. Pony is a more comic-like, often anime/hentai trained finetune of the XL, but it has its own photorealistic checkpoints as well. Illustrious reminds of the art used in manga pocket books art pages or in hentai games and so on. It is the most comic like of the three, but as with the pony, also Illustrious has its own photorealistic checkpoints.

I recommend to download a few checkpoints from all three finetunes, because LORAs are made so that they often work only with the finetune they are made for. SD15 checkpoints are about 2 gigabytes each, and SDXL checkpoints (all finetunes) are about 6-7 gigabytes each.

About Checkpoints in Flux or HiDream

I cannot really tell anything yet, I haven't ever ran any HiDream generations as I am limited to 8Gb of VRAM myself, and my adventures in Flux have been elementary. Someone else can surely enlighten about these.

- N