Stable Diffusion, NovelAI, Machine Learning Art - AI art generation discussion and image dump

[Abeezo]

misc stuff becuz i haven't posted in a while

 

[Catgirls are Love]

Can I use this to make realistic images like the one of Trump getting arrested? I see Midjourney costs shekels now and also requires >discord so fudge that.

Yes, though it definitely helps if you also know how to use Photoshop.

 

[Akashic Retard]

ChatGPT continues to amaze me despite the devs' efforts to reign it in. I had a random thought and wanted to see if it was something other people had thought of already so I could read more about it, but wasnt really sure how to go about googling it. I asked ChatGPT and it gave me exactly what I was looking for. Add teacher to the very long list of roles AI can fulfill/replace.

 

[King Dead]

Yes, though it definitely helps if you also know how to use Photoshop.

It's funny how much seething there's been from "artists" (read: a lot of people drawing porn) about how people are using AI to pump out images in a fraction of the time, when the reality is that most of the really good AI pictures still take a pretty decent amount of time to produce. It's a different skillset and probably not as hard to learn as actually learning to draw, but it's not as simple as just type prompt and get booba.

Well, okay, you can do that and you'd probably get something you could fap to in a pinch. But for quality imagery, it's more involved.

I talked about the process I go through, and there are various things I've had to learn to produce images I'm happy with. Prompt building is a big one, figuring out how to get the AI to understand what I'm looking for and produce something close. I'll often have to render dozens of txt2img results to find one that I think I can work with, and that takes time, especially with my somewhat older card. Then img2img upscaling to refine the base image, often needing to do that a few dozen times as well, followed by inpainting and sketch inpainting to fix smaller issues. If needed, I'll fire up GIMP to fix details that the AI's not quite getting right. And I have to repeat this process with every upscale, slowly but surely working towards an image that I'd call finished.

I'd say that in total, a single hi-rez, highly detailed image can take me a few hours to complete; much of that is rendering time, but it's never as simple as typing a few words and I'm done. I won't claim that what I do is art, but I won't deny that there's a creative drive behind it, and that makes it mean something to me.

 

[Pixy]

NovelAI updated their image generation UI.

The most significant development of all this is that it supports gems.


So where are all the AI-generated cobsons? I've never seen any, but if they're using it to show off their new features, that means someone's gotta be running hundreds of generations of it.

 

[Epic Fail Man]

"Final Fantasy IX body positive"

 

[Post Reply]

100% agree on getting a used Nvidia GPU. I'm planning on using my tax return to snag a second hand V100 for some more high-end oomph.

My budget for a spare GPU was a little tighter, so I picked up a cheap EVGA b-stock 2060 12gb. It's an order of magnitude faster at Stable Diffusion than my 5700xt, but I do bump up against memory constraints when generating higher res images.

 

[King Dead]

My budget for a spare GPU was a little tighter, so I picked up a cheap EVGA b-stock 2060 12gb. It's an order of magnitude faster at Stable Diffusion than my 5700xt, but I do bump up against memory constraints when generating higher res images.

That's one thing that you'll have to learn through trial and error, the limitations of your specific graphics card. One thing that'll help is turning off live previews if you have them on; while it's neat to watch the process from start to finish sometimes, it adds to the memory usage, especially at higher resolutions. I was having my card crap out on me and force a reboot until I figured out I needed to turn it off.

Another is batch sizes. If you're going to run a lot of images (and you will, most times), you can lower the memory usage by generating more batches of smaller size. The size of the batch affects how much memory is needed, but the number of batches doesn't. So for example, five batches of three images will require less memory than three batches of five.

With my 1070, I can do five batches of five at 576x576 and 768x768, but I drop down to three batches of three when I go up to 1024x1024. At 1536x1536, I can only do batch size of 1, maybe 2 before it craps out, so I may do four single batches or two doubles for comparison's sake. Keep messing around with the numbers and you'll figure out what your card can do.

 

[Catgirls are Love]

That's one thing that you'll have to learn through trial and error, the limitations of your specific graphics card. One thing that'll help is turning off live previews if you have them on; while it's neat to watch the process from start to finish sometimes, it adds to the memory usage, especially at higher resolutions. I was having my card crap out on me and force a reboot until I figured out I needed to turn it off.

Another is batch sizes. If you're going to run a lot of images (and you will, most times), you can lower the memory usage by generating more batches of smaller size. The size of the batch affects how much memory is needed, but the number of batches doesn't. So for example, five batches of three images will require less memory than three batches of five.

With my 1070, I can do five batches of five at 576x576 and 768x768, but I drop down to three batches of three when I go up to 1024x1024. At 1536x1536, I can only do batch size of 1, maybe 2 before it craps out, so I may do four single batches or two doubles for comparison's sake. Keep messing around with the numbers and you'll figure out what your card can do.

I've got a 1080, and my general rule is for anything up to 768 batch sizes of 2 or 4 depending on how fast I want results from the prompt, but for anything over 1024 (hi res fix) batch size of 1 even if I might be able to get away with a larger batch size (install xformers; it reduces memory usage by a not insignificant amount).

 

[King Dead]

I've got a 1080, and my general rule is for anything up to 768 batch sizes of 2 or 4 depending on how fast I want results from the prompt, but for anything over 1024 (hi res fix) batch size of 1 even if I might be able to get away with a larger batch size (install xformers; it reduces memory usage by a not insignificant amount).

Ah yeah, I forgot about that, it's definitely a major help. And since there are prebuilt files out there, it's a no-brainer to grab it and install for the performance boost.

I tend towards doing small batches while I'm refining my prompt, then I jump up to my 5x5 when I think I've got it down. I like to have a lot of pics to compare at once rather than generate a few, see if I like any, then generate a few more. Usually it takes a few dozen before I've got a starting point I think I can work with.

 

[Psicopax]

Just started with this around a week ago. Here are 2 images i'm working on atm

 

[Slime Tube]

Decided to try creating some images with "Terran Trade Authority" in the prompt. The results were a pleasant surprise. Hope some of the rest of you enjoy these.

 

[Stephanie Bustcakes]

Decided to try creating some images with "Terran Trade Authority" in the prompt. The results were a pleasant surprise. Hope some of the rest of you enjoy these.

View attachment 4978125View attachment 4978129

The second one is very vaporwave.

 

[Monkey Pink]

 

[Psicopax]

figured I’d post a small progress timeline.

It can be easy at times to generate good images if you are not looking for something too specific, but if you have a scene in mind, you’ll quickly run into issues.

I talked to a friend who said he gets discouraged because most people only upload the best pictures they get from batches of generations and in painting and fixes, and when you can’t replicate good results, it’s easy to feel a little down. So I started saving iterations of images to show him the journey the image makes. Thought maybe it could be fun to post it here too.

So I want to show what I started with and what I ended up with by using a little photoshop to move the rifle stock, in painting and img2img upscaling

Iteration 1 - Generation


Iteration 2 - Photoshop gun stock + Inpaint face.


Iteration 3 - img2img upscale + Gloves added in photoshop (fuck AI hands) + Window and gun smoke inpaint.

 

[King Dead]

figured I’d post a small progress timeline.

It can be easy at times to generate good images if you are not looking for something too specific, but if you have a scene in mind, you’ll quickly run into issues.

I talked to a friend who said he gets discouraged because most people only upload the best pictures they get from batches of generations and in painting and fixes, and when you can’t replicate good results, it’s easy to feel a little down. So I started saving iterations of images to show him the journey the image makes. Thought maybe it could be fun to post it here too.

So I want to show what I started with and what I ended up with by using a little photoshop to move the rifle stock, in painting and img2img upscaling

Iteration 1 - Generation
View attachment 4979910

Iteration 2 - Photoshop gun stock + Inpaint face.
View attachment 4979922

Iteration 3 - img2img upscale + Gloves added in photoshop (fuck AI hands) + Window and gun smoke inpaint.
View attachment 4979930

Thanks for posting these, hopefully it helps people from being discouraged about why their results don't match others'. Like I said, it's rare that you can just get something great from a single txt2img, but the more effort you put in, the better your final result will be. I'll probably do something similar to illustrate my process when I get a chance.

 

[Pendulous Dong]

The thing I find great about this technology is that you can intersperse AI images with real ones, and if the first couple of images are real then nobody will think to look for the fakes.
This shit was done in twenty minutes, imagine what someone with an agenda to push and the will to make it look convincing could do.
Nothing is ever going to be real again.

 

[The Mass Shooter Ron Soye]

View attachment 4985368
The thing I find great about this technology is that you can intersperse AI images with real ones, and if the first couple of images are real then nobody will think to look for the fakes.
This shit was done in twenty minutes, imagine what someone with an agenda to push and the will to make it look convincing could do.
Nothing is ever going to be real again.

Careful, the press might do a hit piece on you.

Good job though.

 

[Psicopax]

View attachment 4985368
The thing I find great about this technology is that you can intersperse AI images with real ones, and if the first couple of images are real then nobody will think to look for the fakes.
This shit was done in twenty minutes, imagine what someone with an agenda to push and the will to make it look convincing could do.
Nothing is ever going to be real again.

Not sure if this has been touched on before, but if you really wanna fuck with people, even those who know stuff about AI, you gotta do one very simple thing.

Ethics in AI generation, and how to ignore it
Here is the thing. If you are using Stable Diffusion, it does something it doesnt really tell you. It adds a watermark that you cant see.
But people looking at the noise and some specific frequencies will be able to see it. I figure its to make life easier for fact checkers.
So lets make it difficult again.
Go into Settings, then to Saving images/grids in the sidebar.
Then scroll almost all the way down and click this handy little box

Be warned tho. Apparently this is very "unethical"

 

[AmpleApricots]

View attachment 4985368
The thing I find great about this technology is that you can intersperse AI images with real ones, and if the first couple of images are real then nobody will think to look for the fakes.
This shit was done in twenty minutes, imagine what someone with an agenda to push and the will to make it look convincing could do.
Nothing is ever going to be real again.

His suit isn't the same in the generated ones but you're right, for the average A&H enjoyer, that's way to subtle to be ever picked up upon. Also these are very very good, sure they're AI generated? I genuinely wouldn't be able to tell.

They need to open the floodgates on this so people know not to blindly believe everything, it's that simple. Any gatekeeping in that regard will just empower the disinfo people and also just delay the inevitable.

Huh, it's almost like they don't want people to form their own opinions, isn't it