AI generator - i am lost

[tanstaafl]

Very insightful.

Flux. I read their page and they seem to have to versions. One for smaller VRam (upto 16GB) and then something upto 32GB.
So that would be obviously a little too much for my system.
Are some of these things like checkpoints actually payware? Just wondering. I wasn't sure about Flux but i think i need a better machine for that.

I tried also webUI. Simpler.

Pony. I read about it but it seems what i came up with, that is mainly meant to be for anime. Or at least thats the main appeal.

The checkpoints. I wonder if you could train it yourself or do you have (seriously) a more mathematical understanding to do that.
I compare it to DAZ or anything 3D where mostly the fake is preferred. I mean, its fantasy anyway but i tend to like more older women as i am as old as them. I appreciate their wrinkles and how age changes your appeal.
At the moment, i am going wild.

I did some interesting images by asking stuff like cat with skittles or what does a dement person see. I think it doesn't really understand and creates things from that sentence.
Still, fun and scary with what it came up with.

Its more artistic and i suppose it can create something out of what ever you tell it.
Are there are more sophisticated prompts?

Oh another point is the resolution. My impression is, that it's small for a reason and you upscale images to the desired resolution. Is that correct? Most that i came across make images 512x512 or 768x512 but not really larger.

View attachment 4833531 View attachment 4833538 View attachment 4833539 View attachment 4833540

For upscaler in comfy (if you're using comfy) open the manger.

Select Custom Nodes Manager and search for Comfyroll Studio and install it.
Then in Model Manger search for RealESRGAN x4 and 4x-UltraSharp and install those.

In your workflow add a CR Upscale Node and Select one of the upscale models you just installed. Test out both see which one you like.

 

[coffeeaddicted]

For upscaler in comfy (if you're using comfy) open the manger. View attachment 4833572

Select Custom Nodes Manager and search for Comfyroll Studio and install it.
Then in Model Manger search for RealESRGAN x4 and 4x-UltraSharp and install those.

In your workflow add a CR Upscale Node and Select one of the upscale models you just installed. Test out both see which one you like.
View attachment 4833584

At the moment i go with both (Comfy and WebUI). The node arrangement is a little harder. I seen some flowcharts with skins and checkpoints and much more of course.
I will try that. So my assumption was correct?

Thank you

 

[tanstaafl]

At the moment i go with both (Comfy and WebUI). The node arrangement is a little harder. I seen some flowcharts with skins and checkpoints and much more of course.
I will try that. So my assumption was correct?

Thank you

Yeah, your assumption was correct. But the sizes you mentioned are generally SD1 sizes. For SDXL most people use 832x1216 or reversed for horizontal.

 

[coffeeaddicted]

Yeah, your assumption was correct. But the sizes you mentioned are generally SD1 sizes. For SDXL most people use 832x1216 or reversed for horizontal.

Ah, really?
I always look at the description and the bad part is, it isn't always mentioned or maybe its somewhere else.
Never stop learning. lol

 

[Dark Lord Nova]

I decided to do a clean installation of the standalone version of ComfyUI directly from their site—no Pinokio this time. I'm still keeping SwarmUI as my backup generator, but I really want to dive back into ComfyUI and use it properly. Honestly, I haven't touched it in a while except that sort of built-in version of it in SwarmUI that's limited. When I first tried it, I was still figuring things out and never really got the hang of it. Now I'm ready to explore it seriously. What custom nodes do you actually recommend?

 

[tanstaafl]

I decided to do a clean installation of the standalone version of ComfyUI directly from their site—no Pinokio this time. I'm still keeping SwarmUI as my backup generator, but I really want to dive back into ComfyUI and use it properly. Honestly, I haven't touched it in a while except that sort of built-in version of it in SwarmUI that's limited. When I first tried it, I was still figuring things out and never really got the hang of it. Now I'm ready to explore it seriously. What custom nodes do you actually recommend?

The only once that I install by default are ComfyUI Impact Pack, Comfyui Impact Subpack, Comfyroll Studio, and, of course, the

You must be registered to see the links

(this one from github, follow install directions on the page).

Edit: I'm also currently working on getting the LLM Polymath node set up in my comfy flow. It allows api calls to LLM (large language model) for better prompt interpretation that should (is supposed to) drastically increase overall results. (requires API access to specific LLMs (not free) or an install of OLLAMA a local LLM.

 

[IBreedWives]

Usually game labeled as real porn use small videos, so photorealistic game that will replace it would focus in animate any generated image but it require too much time compared to static images.

i would much rather have high quality photorealistic static images than shitty low quality porn clips, i think devs are using real porn wrong in games here, a slideshow of porn gallery pics would be better than most of the shitty clips i have seen in real porn games

 

[bigbootyluver69]

I'm not very active on here so I'm a tad slow but to answer your questions:

Very insightful.

Flux. I read their page and they seem to have to versions. One for smaller VRam (upto 16GB) and then something upto 32GB.
So that would be obviously a little too much for my system.
Are some of these things like checkpoints actually payware? Just wondering. I wasn't sure about Flux but i think i need a better machine for that.

Not really, the only time checkpoints are payware are the big ones like OpenAI image gen, google's and a few others as well but otherwise if your machine can handle it you can just download the model and start image generating. I know Flux has a paid verison of their models but the Flux.dev is really good so most people just opt to download it instead of paying for using it on Flux's servers.

Pony. I read about it but it seems what i came up with, that is mainly meant to be for anime. Or at least thats the main appeal.

The checkpoints. I wonder if you could train it yourself or do you have (seriously) a more mathematical understanding to do that.

For the most part yes, pony is mainly used for anime type of image generations but there are some real banger realistic checkpoints that use pony.

The checkpoints. I wonder if you could train it yourself or do you have (seriously) a more mathematical understanding to do that.

From what I've read about training checkpoints is that it requires a lot, like A LOT of GPU power. Most people rent out a couple of A100s to train since it's cheaper than trying to train a checkpoint at home. Also, you can do what's called a checkpoint merge that is like a smaller training session? I can't remember the exact terminology but it allows you to take 2 checkpoints and merge them into 1 new one. I haven't really looked into merging checkpoint tbh. Creating your own LORAs, on the other hand, tends to be faster and more flexible than training checkpoints. LORAs are like smaller version of a checkpoint, it allows a checkpoint to "learn" about a new concept without having to retrain the whole checkpoint.

Its more artistic and i suppose it can create something out of what ever you tell it.
Are there are more sophisticated prompts?

Oh another point is the resolution. My impression is, that it's small for a reason and you upscale images to the desired resolution. Is that correct? Most that i came across make images 512x512 or 768x512 but not really larger.

What might help is you can paste an image you like or want to create into AI, like ChatGPT, Google's Gemma, or use an image tagger like WD14-Tagger and ask it to create a prompt in SDXL to create the image. I personally use WD14 -Tagger since i mainly generate NSFW images and most online AI's will refuse to describe the image lol. This is how I've learn what text invokes what in image generation.

I saw someone else comment about the resolution and yeah, the 512x512 is for SD1.5 since that is the resolution it was trained on. If you're using SDXL checkpoints you'll want to use 1024x1024 or any of the other resolutions for SDXL. If you look up SDXL image resolutions you'll get a list of different sizes for different aspect ratios. Afterwards you'll want to upscale the image, there's a default workflow for latent, and i believe also model upscaling in ComfyUI in the templates. So yes you're right, ideally you start with the small resolution, ex: 1024x1024, then you would use a different workflow to upscale your generated image to X resolution, ex: 2048x2048.

 

[bigbootyluver69]

I decided to do a clean installation of the standalone version of ComfyUI directly from their site—no Pinokio this time. I'm still keeping SwarmUI as my backup generator, but I really want to dive back into ComfyUI and use it properly. Honestly, I haven't touched it in a while except that sort of built-in version of it in SwarmUI that's limited. When I first tried it, I was still figuring things out and never really got the hang of it. Now I'm ready to explore it seriously. What custom nodes do you actually recommend?

Honestly it kinda depends on what you're trying to create but I personally recommended IP-Adapter, Controlnet, Impact pack, ComfyUI Essentials, WD14-Tagger, Inpaint-CropAndSitch, and InstantID. WAS node suite, Rgthree, and Easyuse are good nodes that provide utility nodes as well.

ControlNet and Impact pack make up 80% of my workflows because once you understand the basics you can do some cool stuff with them. I haven't gotten IP-Adapter to work the way I want it all the time, but when it does boy it's crazy to see how it can generate another image based on an image you provide it. It helps with character consistency and that's currently my project atm.

 

[Dark Lord Nova]

The only once that I install by default are ComfyUI Impact Pack, Comfyui Impact Subpack, Comfyroll Studio, and, of course, the

You must be registered to see the links

(this one from github, follow install directions on the page).

Edit: I'm also currently working on getting the LLM Polymath node set up in my comfy flow. It allows api calls to LLM (large language model) for better prompt interpretation that should (is supposed to) drastically increase overall results. (requires API access to specific LLMs (not free) or an install of OLLAMA a local LLM.

I just installed ComfyUI straight from its own official site—it gave me three options: Windows and Mac executables, or the portable GitHub version. I went with the Windows installer, and it automatically set up the ComfyUI Impact Pack, Subpack, and Manager for me.

Right away, I noticed this version looks way better—cleaner interface, more intuitive, and it even checks for updates and includes workflow templates. It honestly feels smoother and more polished than the one I used with Pinokio or even the older standalone I had before that. I'm actually really into it now.

Sounds like you grabbed the GitHub version? Not sure if it looks or works the same as mine.

Honestly it kinda depends on what you're trying to create but I personally recommended IP-Adapter, Controlnet, Impact pack, ComfyUI Essentials, WD14-Tagger, Inpaint-CropAndSitch, and InstantID. WAS node suite, Rgthree, and Easyuse are good nodes that provide utility nodes as well.

ControlNet and Impact pack make up 80% of my workflows because once you understand the basics you can do some cool stuff with them. I haven't gotten IP-Adapter to work the way I want it all the time, but when it does boy it's crazy to see how it can generate another image based on an image you provide it. It helps with character consistency and that's currently my project atm.

I will definitely check them out and add those if they give me way better images and video clips. Well I already have some of those installed like Controlnet and Impact pack that came with ComfyUI.

 

[tanstaafl]

Sounds like you grabbed the GitHub version? Not sure if it looks or works the same as mine.

Yeah, I grabbed the github version. Think of it like installing a linux version with fewer packages installed, but lets you pick and choose exactly what you want.