Discussion about A.I.

[fallenmystic]

I'm in the process of refining my RP settings. I've been trying to build a "lore book" using "World Info" SillyTavern provides. Specifically, I made a slave trainer character and associated lores to let him understand what things a new slave must learn and how.

It's been working reasonably well, although the AI sometimes ignores some settings. Below is an example of the lore book about slave rules and how it affected one of my RP sessions with the aforementioned character:




(1. Images are tiny, so I just embedded them in full resolution. 2. Terse sentences in the entry are intentional, to save token count.)

 

[captivecuties]

I'm in the process of refining my RP settings. I've been trying to build a "lore book" using "World Info" SillyTavern provides. Specifically, I made a slave trainer character and associated lores to let him understand what things a new slave must learn and how.

It's been working reasonably well, although the AI sometimes ignores some settings. Below is an example of the lore book about slave rules and how it affected one of my RP sessions with the aforementioned character:

View attachment 1365655

View attachment 1365654

(1. Images are tiny, so I just embedded them in full resolution. 2. Terse sentences in the entry are intentional, to save token count.)

The lore book is very similar to the way I Train My property.
Very cool.

 

[Stungiskhan]

I just realised I forgot to add the actual link. Since nobody complained about it, I suppose it shows almost no one is interested in the subject yet.

Anyway, here's the link to the website mentioned in the previous post: The Local LLM Settings Guide/Rant

In general, I’m not interesting on role-playing but for most of the new technologies and possibilities. AI characters/players have exist in games for long and got better and better. Getting a bit curious which kind of improvents this type of model with very heavy requirements provide.

 

[fallenmystic]

In general, I’m not interesting on role-playing but for most of the new technologies and possibilities. AI characters/players have exist in games for long and got better and better. Getting a bit curious which kind of improvents this type of model with very heavy requirements provide.

I'm also very interested in the subject, making a BDSM sandbox game being my long-term goal in learning AI/3D stuff. We're already seeing quite interesting attempts at using generative AI technology to make more immersive gaming experience:

 

[Stungiskhan]

For me, that sounds interesting. I hate modern scripted games. I also don’t want the game to be too complex, so I need to make it for my second life to be good at it. I could play a game which would be relaxing to play, benefit from some clever thinking and offer a peek to the fantacy world with something more interactive than normal picture sets.

 

[fallenmystic]

I got too impatient to wait for ControlNet support to land in SD.next, so I switched to ComfyUI, a more advanced tool for Stable Diffusion.

I managed to make a decent render with it, utilising ControlNets as I used to with SD.next. I still don't know how to inpaint in a larger resolution than the masked area which is needed for things like depicting brands or adding details to the face.

I'm glad that I switched, however. Even though it's far more complicated to do things in ComfyUI compared with how it was with SD.next, it feels good to have full control over the whole process. The workflow I made for the render I linked above looks like this:


I like how ComfyUI caches the output of each node, which has greatly saved my time spent on finetuning my render. SDXL was particularly slow on SD.next, especially with Loras. I had to wait for a minute or so every time I changed anything in SD.next, but now I can see the result in a few seconds when I change something further down the pipeline.

However, the best feature I found of ComfyUI was how it allowed me to set up a reproducible, non-destructive workflow for each render. It means I can save the whole process of creating an image and loading and updating it anytime I want.

I've been "remastering" my works constantly as I learned something new. Having a way to revisit and modify the way I created each of my renders would greatly help me keep enhancing them.

A quick update on LLM before I go: recently a brand new model called Mistral was released and it has beaten larger models despite being 7B in size. I tested one of the merges (Amethyst-13B-Mistral-GGUF) and was greatly impressed by the result. RP with AI keeps getting better by the day.

 

[fallenmystic]

A few quick observations after having done my first serious render in SDXL/ComfyUI:

Photorealism is far easier to achieve compared with how it was with SD 1.5. What realism-focused checkpoints I've tried so far were all fairly competent in generating photo-like images. However, it'd still be nicer to have top-quality models like epiCRealism/epiCPhotoGasm for SD 1.5.

In contrast, I'm disappointed that those photorealistic checkpoints fail to depict nipples or genitals in sufficient detail. Maybe it can be mitigated with inpainting or using some Loras (which aren't yet available in comparable numbers/variety with those available for SD 1.5). However, having a good base model that can depict those anatomic details well would be better so that it can be merged with other photorealistic models.

I'm still new to ComfyUI so I can't make a definitive judgement. But from what I've tried so far, it's a fantastic tool that can be more efficient and powerful than most other frontends. But there are things that were easy in A1111/SD.Next but proved to be quite a hassle in ComfyUI, like inpainting with "masked only" mode, for example. So, my recommendation for SD frontend is,

• If you only have a casual interest in AI-assisted image rendering, go with one of the free options like AI Horde. Civitai itself offers a temporary free service also, which is awesome since you can try any model hosted on the website.
• If you want to make your art or need more control over the images you generate, go with either A1111, or SD.Next. They are by far the most popular choices among AI image generation enthusiasts, so you can't hardly go wrong with them.
• In case you are familiar with a node-based workflow (e.g. Blender's shader tree, or Substance Designer), or want to have full control over the whole image generation process, try ComfyUI. You can develop your own workflow and reuse it easily while making it more efficient and flexible. It can be a good way to learn how Stable Diffusion works also, if you feel interested in the subject.

 

[Danillegs]

I got reminded of having read that SDXL can render legible texts, unlike its predecessor, so I did a little test:

View attachment 1363279 View attachment 1363280

It looks like now I can finally render tattoos and brands without relying on an external program.

Wonderful, amazing !!

 

[Nyghtfall3D]

I've got a few questions regarding the degree of control AI offers.

Can you make finite adjustments to body parts or facial expressions to create the precise kind of pose you want? Right down to fingers and toes?

Can you create and position lights wherever you need them to create the kind of mood or dramatic tension you want, and adjust their size, color temperature, or intensity?

Can you create and position cameras for a cinematic look?

 

[fallenmystic]

I've got a few questions regarding the degree of control AI offers.

Can you make finite adjustments to body parts or facial expressions to create the precise kind of pose you want? Right down to fingers and toes?

Can you create and position lights wherever you need them to create the kind of mood or dramatic tension you want, and adjust their size, color temperature, or intensity?

Can you create and position cameras for a cinematic look?

First of all, thanks for dropping by! I always respect people who are willing to engage in a rational argument, and I enjoy talking with them.

To answer your question one by one, yes, you have fine control over body parts or facial expressions. The most common method of doing so is using an "openpose" variant of controlnet model. I even made a workflow of automating the generation of the needed images from a Daz3D character imported into Blender:



Another way of doing it is just drawing a rough sketch of the pose by hand. I often draw a crude image in Krita and develop my render based on it,

As for the lighting, you'll need a different - but also more intuitive approach - than what is usually employed by 3D artists. Remember the whole concept of PBR is an attempt to mimic how light acts physically, which is both the strength and limitation of the traditional 3D rendering workflow.

While PBR helped create highly realistic images, it often falls short of achieving true photorealism. It's mainly because it's impossible to accurately simulate every aspect of light particles hitting surfaces in the real world, so we have to simplify things to make it possible to render an image on a typical desktop PC that way. (And that's why most of the Daz3D renders feature plastic-looking skins, despite using gigantic 4k-8k textures).

The concepts you cited, like adjusting the temperature/size/intensity of a light source are needed for the traditional physics-based approach. With AI, however, if you want a cinematic look, you can just find good references (e.g. style Loras like this, or these) and mix & match them to create the look you want.

I admit that sometimes you just want to put a spotlight in an exact place, for example, which is doable but quite difficult to achieve with AI. But let's be real. How many Daz3D users do you think would be able to replicate the lighting used (which is a pretty simple one, by the way) for the examples in this Lora, for instance?

And if we are to list things that AI can or cannot do, how about what traditional 3D applications like Daz3D can do? While I personally feel confident that I'd be able to mimic at least most of the Daz3D renderings posted on CF in either Daz3D, Blender, or with AI, what percentage or these random AI renders you see on this page that you feel a skilled Daz3D user can mimic?

And I'm not even talking about photorealism but the freedom of expression here. Can you, for example, depict a mechanical snail made of spaghetti noodles in Daz3D, like one you see in this example? Or can you render your Genesis Female character in the style of ancient Chinese ink painting like I tried here?

Compared with AI, traditional 3D design tools provide much less freedom of expression to their users, and Daz3D is even more limited in this regard, considering how it's made for those who don't have much 3D skills to help them render scenes using ready-made assets without much hassle.

If you still want to see how many "knobs" AI provide that you can tweak to your heart's content, I can show you things like what workflow I used for my latest render, like this one:



But to me, what is important in this matter is not how many controls such a program provides to its users but the degree of freedom it allows them to express what they have in their minds. In that regard, I feel AI to be a much superior tool to anything we had before, which can be a blessing rather than a curse to any artist who wants to explore the freedom of creativity it gives them.

 

[Nyghtfall3D]

And I'm not even talking about photorealism but the freedom of expression here. Can you, for example, depict a mechanical snail made of spaghetti noodles in Daz3D, like one you see in this example? Or can you render your Genesis Female character in the style of ancient Chinese ink painting like I tried here?

I got into 3D art because I want to express my fantasy death fetish fantasies, and fetish videos remain prohibitively expensive to finance. I have no interest in rendering mechanical snails or replicating someone else's artistic style. I want to and enjoy creating fetish-themed 3D art using pre-fabricated human figures, clothing, environments, and props. I want the freedom of navigating a virtual environment and setting up scenes with my own eye for a cinematic presentation. I get an enormous rush out of watching my pain-staking attention to detail come to life in the final render, especially the more lifelike a pose looks. For example, I rendered the attached image below in April, and I'm still stoked by how well I did making my protagonist brace herself against the sink while trying to keep from being drowned.

Granted, apps like Daz and Poser have their limitations, but they still offer a greater degree of control than what I understand AI is capable of.

To answer your question one by one, yes, you have fine control over body parts or facial expressions.

Can you control which direction a character's eyes are looking, or the position of their eyebrows?

 

[fallenmystic]

I have no interest in rendering mechanical snails or replicating someone else's artistic style. I want to and enjoy creating fetish-themed 3D art using pre-fabricated human figures, clothing, environments, and props.

I think we better start by reiterating our respective positions on the matter before we start. As I made it clear in the other thread, I'm absolutely fine if you feel content with Daz3D as a tool to visualise your artistic vision. I even encouraged you to keep using it to create your art.

On the other hand, you've been arguing against the feasibility of AI being a legitimate tool of art. So, the case with the mechanical snail was just an example to show you how AI allows much greater freedom in style and subject matter compared with Daz3D.

In this regard, whether or not you personally feel content to use what limited selection of models/styles Daz3D provides doesn't matter. In this case, what is important is how you'll defend your position that AI isn't even a legitimate tool of art since it lacks "human creativity" when it, in fact, allows far more diverse options for expressing the artist's creativity than Daz3D does.

Again, I don't hate Daz3D or think it's not a valid tool for creating serious artwork. To me, whatever tool that allows people to express their artistic visions in ways they want is a perfectly legitimate method to create art.

But I cannot help but feel ironic when Daz3D - a program made to allow those without 3D skills to make decent renders by purchasing and arranging ready-made assets - users argue how AI cannot be a valid way of creating art citing such reasons as "anybody without skills can use it" or how it "deprives the freedom to express artistic vision", etc. when I regularly use "real" 3D tools like Blender or SubstancePainter as a part of my AI workflow.

I want the freedom of navigating a virtual environment and setting up scenes with my own eye for a cinematic presentation. I get an enormous rush out of watching my pain-staking attention to detail come to life in the final render, especially the more lifelike a pose looks.

Preferring a tool that allows free navigation in a virtual space is something which I totally understand and respect. It's one of the reasons why I like to start my AI render from a Blender scene. I'll just add that what you described of the feeling you have when you see the final render isn't exclusive to Daz3D users.

Granted, apps like Daz and Poser have their limitations, but they still offer a greater degree of control than what I understand AI is capable of.

I believe you feel this way only because you exclude those cannot be done with Daz3D as something inherently impossible to achieve. Take an example I mentioned in my previous post, like how AI allows me to render a slave girl in the style of traditional Chinese ink art.

Is there any reason other than it's something Daz3D cannot do that you exclude it in your criteria of "degree of control/freedom of expression"? Ability to choose in what style you can make your render is still a matter of having artistic control, isn't it?

Can you control which direction a character's eyes are looking, or the position of their eyebrows?

Sometimes it's tricky but entirely possible. There are multiple ways of achieving it too. I usually just draw it on a ControlNet image by hand in such cases. But if you want total control over facial expressions, you can use something like this too:

 

[Stungiskhan]

I don’t think it’s worth of trouble to argue about these two methods because how I see it, both are still in very early stage of development and could benefit from ‘competition’. Personally I’d like to see a combination of both to enable much more realistic and easier way to generate art, or whatever we want to call it.

I think AI has a lot to contribute on this, while I’m not a biggest fan of current stable diffusion type of approach, which aims directly on the end result.

I have much more experience in 3D modeling, so I’ll point out some problems I have faced and think needs improvement.

First, ready made models are imperfect and usually cost lots of money or it takes days or weeks worth of work to make one of your own. While the shape can be altered, skin details are fixed on the available characters, and they are very important. As we get closer to photorealism, making a good skin becomes exponentially more difficult.

While some DAZ models have very realistic face in right light, the rest of the body is far from it. Skin is typically painted in a rest position, making it look very unnatural when being stretched from the joints. Details should be added in a stretch position, but making nude figure to look realistic, is not desired priority. While there’s some joint fixing morps in DAZ characters, classic bent over, legs spread position looks terrible, with unnaturally folding and stretching skin and lack of details.

I use blender, because it allows fixing those problems, but it’s days worth of work for one model, and still the result is somewhat unnatural, unless using some high resolution photo references or live model, which are typically not available. This could easily get much worse if stepping further from current technology, which have quite many flaws.

Many people think that current raytracing is physically accurate. In reality, it only use physics to mimic some very basic phenomenas. As a result, models look good only in very specific lighting conditions related to shaders their skin uses. It has great limitations on the freedom to use lights.

The basic problem is to make the skin surface have enough details to highlight the shape and expressions but still look realistic, or at least good. In many so called hyper realistic models, skin details are highly exaggerated to create such effect.

Finding good lighting is difficult and non intuitive and dependent on the renderer. That’s because of the limits in render technology. Typically renderer models diffused reflection, which reflects diffuce color into all directions same way. They also model light penetration through the skin, but by very simple approximation. Then they model different types of specular reflections, like specular, metallic or clear coat, which typically have roughness value, indicating how sharp the reflection is. It is meant to mimic microscopic surface properties, finer than details in a texture, dispersing reflected light into different directions.

In practice, this roughness value is not constant for certain spot in a skin but varies based on light direction, because of microscopic pores, layers, grease or such. That makes all important details step up better in practice than in 3D scene. Light penetration modeling, called subsurface scattering, have a tendency to blur the details further, because of simplicity of modeling.

Modeling everything to overcome these problems would be very difficult and also models would become huge. AI could be used to generate textures and shapes for current rendering models and tweak textures for some specific situation and lighting to provide desired results. It would be easy to add dirt, scratches, dimples or pores to be visible in desired way.

AI could also create facial expressions and poses, but with 3D modeling, giving artist a possibility to tweak everything and provide repeatability for different shots.

 

[fallenmystic]

I don’t think it’s worth of trouble to argue about these two methods because how I see it, both are still in very early stage of development and could benefit from ‘competition’. Personally I’d like to see a combination of both to enable much more realistic and easier way to generate art, or whatever we want to call it....

That almost perfectly mirrors what I feel about the subject. Even though I primarily use AI to create my renders nowadays, I still use Blender, and I acknowledge there are many use cases where the traditional approach is more suitable, like in creating video games, for example.

Also, I've spent quite an effort to approach photorealism in Blender until I hit a similar wall you described in your post. I tried various approaches including drawing skin textures in Substance Painter, purchasing high-quality photo scans, and so on. But as you said, all those options required too much time and money to be really feasible for a hobbyist like me, especially when considering all those efforts need to be repeated for each character, in case you want to achieve true photorealism (because skin details cannot be the same between people, like you pointed out.)

And my conclusion from the experience was also pretty much the same as yours. While the traditional 3D workflow still has and will retain its uses, there are a lot of things that can better be done using AI, like generating unique skin textures, fixing the topology of a 3D scanned character, making UV maps, skinning the rig, generating animations out of videos, and so on.

It'll also benefit Daz3D artists as well since things like generating a photorealistic skin pack for the base character have remained out of the reach of ordinary users so far, who aren't highly skilled in 3D modelling or don't have sufficient time and money to burn. It'll be fantastic if you can throw some high-resolution reference photos of a person and AI generates a fully working character morphs and a texture pack out of them.

 

[Anna98]

BTW even all modern denoisers like OpenImageDenoise are AI based. So at some point edge between "AI art" and "AI enhancement" may become quite fuzzy.

 

[Stungiskhan]

BTW even all modern denoisers like OpenImageDenoise are AI based. So at some point edge between "AI art" and "AI enhancement" may become quite fuzzy.

For sure, and AI in some forms can soon be found practically from everywhere.

For my opinion, AI should be used as a human tool, to make easier to create something or open new possibilities. Using AI to entirely replace human in a process is not good idea and often don't provide good results.

Current AI based image creation systems can be both. Complete solution for beginners and tool for professionals. It's good to watch where the development is going. It's good to have dedicated people who takes it seriously and stretch the limits.

There's always a risk that models become so complex and heavy so they run only on commercial servers and only few people are responsible of developing them. Others have to pay from usage and use what is offered.

Making any type of picture with all possible styles and variations is a complex goal. Dividing it into smaller peaces, like in aided 3d model creation process, enables variations and leave room to creativity.

 

[fallenmystic]

Making any type of picture with all possible styles and variations is a complex goal. Dividing it into smaller peaces, like in aided 3d model creation process, enables variations and leave room to creativity.

I agree with what you said except for the above, which I think can be rather different in practice. AI models need training data, and it can be difficult for individual hobbyists to make something like a dataset of high-resolution full-body scans, which would be needed to make an AI that generates photorealistic skin texture sets.

In comparison, everybody can take digital photographs nowadays, and you can train a decent model for a specific concept or style using just 20+ photos. That's how we have a repository for AI like Civitai where you can freely download over 166k+ models, which is quite impressive considering Stable Diffusion has been around for barely a year now.

 

[Stungiskhan]

I agree with what you said except for the above, which I think can be rather different in practice. AI models need training data, and it can be difficult for individual hobbyists to make something like a dataset of high-resolution full-body scans, which would be needed to make an AI that generates photorealistic skin texture sets.

In comparison, everybody can take digital photographs nowadays, and you can train a decent model for a specific concept or style using just 20+ photos. That's how we have a repository for AI like Civitai where you can freely download over 166k+ models, which is quite impressive considering Stable Diffusion has been around for barely a year now.

My vision goes little further. It includes AI, which can create a body texture and maps for wide variety of conditions. It could use completely different method than current diffusion.

It could know hundreds of different skin types and how they behave in certain light. It could know which areas of the body have certain skin and details and combine those based on users request. There could also be pre-render optimizing pass to modify and improve textures in pre-render phase using object normal and light position information. That could allow more realistic or more artistic renderings without heavy modeling and excessive calculations.

Taking photographs is easy but taking good and artistic photographs is difficult. Also in reality, lighting is very important but much more difficult to setup than in digital world. I have many times chased a perfect light when taking pictures, but it hardly never matches with situation to have something desirable in that picture.

In reality, dark room and one candel can make a nice picture of human model but in renderings, very unlikely. The main reason is imperfect modeling of surfaces and lights, but AI can help a lot.

 

[fallenmystic]

I started working on a new scene, and I'm also experimenting with various new approaches to make constructing complex scenes easier.

I remember some other member who dismissed AI image generation as a toy once said he didn't understand why I usually spend several hours for creating a render, assuming there was little more than prompting in the process. And here's why :



That was the ComfyUI workflow I created for the current render project, which is still WIP. The experiment has been only half successful though, as I'm still struggling to fight some of the long-standing issues with making complex scenes.

ComfyUI, however, gave me enough new tools to try different approaches so I'm fairly certain that I'll be able to learn enough things to improve my renders.

By the way, is there anyone interested in the technical details of such a workflow? I'm going to share the workflow and input images once I'm done so that others can get ideas from and experiment with them. But if there's enough interest, I might try to explain what each stage does and why I constructed it that way.

 

[fallenmystic]

As I mentioned above, I'm working on a new scene, the result and workflow I will share here once it's done.

I still need a couple of days to finish it, but since we have the 5 images per thread per day rule in effect, I decided to post a part of it now to share more later.

The idea came from a discussion I had with Zardoz here. As such, I decided to make a whipping scene by utilising my WIP Blender character and making it an opportunity to learn ComfyUI.

Rendering a Whipping Scene Using AI​

Part 1 - Preparation​

The character in question is made in Daz3D first and imported to Blender. Then I added additional things like restraints, brands, and a special setting which automatically generates AI ControlMap images. As I mentioned earlier, you can incorporate various traditional design tools like Daz3D, Blender, and Krita into your AI workflow to great effect, and this could be a good example of that.



I made an asset library for rendering slavery images, as you can see above, by the way. It's interesting to note that the characters are from 3 different sources: Daz3D, HumGen3D, and Procedural Crowd. As I'm going to apply the depth of field effect on the camera, and because this will be just a base image for AI, I won't have to use a high-poly, large texture model for every character.

The finished render is posted below:



I would have thought it was decent and posted it as a final render if it was before I learned how to use AI. Now we can go a step further and make it look photorealistic as well, the result and workflow of which I'll share in my following posts.

As I mentioned earlier, this setup also automatically generates various ControlMap images like this:



I only saved this OpenPose image because the scene is relatively "simply" so other maps like depth or lineart can be more easily generated within the AI tool itself.