AI generated fetish images

[dhf7b8g]

What did you buy? I'm looking now at 4070, which is to be released in April. Nvidias with 12GB  are expensive ... 😕

I went big and bought a 4090, its ridiculously fast and completely unnecessary. I wouldn't recommend it as a purchase to anyone unless you want to have the best of the best regardless of other factors.
I had an early model 3080 before so I wasn't really lacking in graphics power but the 10GB of VRAM was giving me a few issues even outside of AI generation. It still goes for quite a bit on ebay (only £70-80 less than what I paid for new two years ago!!) so I plan to put it up later this month or early march to offset some of the cost of the 4090.

Just as a speed comparison my 3080 would generate at ~15 it/s on DDIM 512x512 which was aeady really fast but the 4090 achieves ~32 it/s in single images and a technical ~48 it/s with a batch size of 4. The 24GB of VRAM lets me really push the resolution too without any performance degrading options.
Currently upgrading my PC into the "dream rig" but the final piece of the puzzle hasn't came out yet...

[dhf7b8g]

In my opinion the "openpose" detection is the most interesting and useful for generating realism with depth map being a close second. It allows you much more natural (or kinky) poses in generation without impacting the quality through a different method like basing a generation on img2img. You could even take a quick phone photo of yourself in the perfect pose you want to generate!

Here are a few non cherry picked openpose examples I generated on my own machine using images I found on google and this webui extension. The process goes as follows:
The input image is a pretty simple model pose:
   
The OpenPose model was able to create this pose map:
   
And finally SD was able to generate this image:
   


Overall generation quality still relies heavily on your model but gone are the days of trying to intricately craft a pose through a prompt. The only descriptor I gave in the prompt was "1girl" the rest was my cookie cutter base prompt with the quality keywords.
Seems ideal for getting tricky or unique bondage poses into generations without needing to spend time training that concept into the model

[Like Ra]

I went big and bought a 4090

YEAH!!! 👍 COOL!!!!!

Currently upgrading my PC into the "dream rig" but the final piece of the puzzle hasn't came out yet...

More info, please!

And finally SD was able to generate this image:

... whoa ... !!!!



And please keep the "how tos" coming!!!!

[dhf7b8g]

More info, please!

My current specs are:
Ryzen 9 3900x
32gb 3600 cl16
Nvidia RTX 4090
(I also have a spare "burner" AMD GPU for virtual machines and hackintosh, but that's not always running)

I've had some worries about my CPU performance for a while, specifically the single core performance, which was notoriously mediocre on that line of AMD chips. The arrival of the 4090 set it in stone. Even with a +500MHz all core overclock, the CPU just isn't fast enough to keep up, although it can still brute force its way through multicore workloads.
Initially I was thinking of buying a 5800x3d, stupid fast chip and cheap platform cost as I'm aeady on AM4, but I didn't want to lose cores even if it just was 4, and I wouldn't have any upgrade path in the future. Lucky for me and less so for my wallet AMD are releasing their AM5 x3d CPUs at the end of this month with higher end variants than a 800x part. They are releasing 900x and 950x variants which are both 12 and 16 core parts respectively. It does mean I need to buy a new motherboard and ram along with the CPU though...

The "dream rig" will be
Ryzen 9 7950x3d
64gb 6000 cl30
Nvidia RTX 4090

Right now, I can't think of better hardware or a use case I have which won't run on this PC, which is why I see it as the dream rig. I just need to patiently wait until launch day, so I can pick one up. I just hope CPU power requirements aren't too high, I've aeady seen my GPU alone draw 550W ❗ 

And please keep the "how tos" coming!!!!

Yeah, for sure I'll keep doing them when I find something interesting, as I enjoy exploring the bleeding edge anyway. I think they are closer to showcases rather than how tos although if anyone curious or stuck does need some guidance I would be happy to help so don't hesitate to ask.

Not too sure what to do next... You (or anyone else) have anything you would like to see? Could always do a realism tips and tricks post or do some training, although I'm not sure what to train either  I'm completely open to ideas.


Also about your 4070. Seems like a reasonable choice. Hard to say where it would land on a benchmark chart but I'm confident it would be able to comfortably generate and even train with optimizations with its 12gb (optimized repos are now just above 10GB which is great although the reduced memory usage does come at a performance cost). A used 3090 is always an option too if you really want that extra VRAM, they are still the go to choice for training. They pretty consistently go for £200 less than what I could buy a 4070ti for which is about where I would expect the 4070 to be priced.

[Like Ra]

nitially I was thinking of buying a 5800x3d, stupid fast chip and cheap platform cost as I'm aeady on AM4

That's why I bought a 5950X - second-hand, but sealed new. 16 cores, 5.2GHz.

I've aeady seen my GPU alone draw 550W

What is your PSU?

BTW, are you on Windows?

[dhf7b8g]

What is your PSU?

BTW, are you on Windows?

Corsair rm1000x so I have leeway as long as I don't have the AMD GPU going at the same time. OCP almost certainly wont trigger unless I get both GPUs and my CPU to spike at the same time but I would rather not risk that

I am not. Arch linux is my "default" operating system which I do almost everything on including AI but I do have a windows install for when I do windows based software dev or play games.

[Like Ra]

Corsair rm1000x

This one is a cool & quiet! I'm trying to choose between rm850x and the latest be quiet! Dark Power 13 850W. The latter is Titanium certified, and despite it does not have the fan-stop mode, it's almost absolutely quiet. So, I'm leaning towards the be quiet! one.

Arch linux

Oh, cool! I'm on Ubuntu. So, whatever works for you, should theoretically work for me (VRAM aside...)

[dhf7b8g]

This one is a cool & quiet! I'm trying to choose between rm850x and the latest be quiet! Dark Power 13 850W. The latter is Titanium certified, and despite it does not have the fan-stop mode, it's almost absolutely quiet. So, I'm leaning towards the be quiet! one.

I would pick the be quiet one as well

Oh, cool! I'm on Ubuntu. So, whatever works for you, should theoretically work for me (VRAM aside...)

Yep, linux gets most of the bleeding edge advancements/features first too so we are at an advantage 😊 I even run ~10% faster in generations than my friend with a 4090 and a faster cpu on windows although 29it/s to 32it/s isn't anything to write home about

[dhf7b8g]

I was able to train the concept of a female chastity belt in only 68 seconds!!!

This is ridiculously easy and I don't know what to do with all the possibilities.
There is a new training method called LoRa which works similarly to overall model training but on a much faster and much smaller scale. It outputs <200MB "extensions" you can run alongside any model to add the trained concepts. File size varies based on training settings but you can train a concept into 10-15MB if you want to. The GPU requirements are reasonably low too with 8GB being the lower limit

It took a total of 5 minutes to complete all steps and the training. All I had to do was download 20 reasonable quality images of what I want to train, caption these images (I used an automatic captioning model and manually added the "chastitybelt" trigger word) and just click train

I trained using very generic settings and no regularization images which is the easiest way to do it but with some tweaking I think I could get some even more impressive results

I attached a few images I generated using this newly trained LoRa.

I especially like how creative it was with this one as there was nothing like it in the training data
   

[Like Ra]

There is a new training method called LoRa

Links? Tutorials?

I used an automatic captioning model and manually added the "chastitybelt" trigger word

What is it? Links? Tutorials? 😉

I attached a few images I generated using this newly trained LoRa.

Simply amazing! I wonder about the tags "erection", "bulge" and chastity cage" 😋