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

  • 🐕 I am attempting to get the site runnning as fast as possible. If you are experiencing slow page load times, please report it.
Train Operator said:
took literally 6hrs but here's my results

View attachment 6109978
Click to expand...
This is if anything even more technically polished than the previous one. The level of realism as it pans around her is astonishing. It really looks 3D and yet the parallax behind her makes the perspective both strange and neat at the same time. It's almost what I might term hyper-real.

Though I have to say, I really like the shock value of the previous one! :D

Now, I said I'd report back about ROCm 6.1.3 and its WSL support. I'm happy to say that with a minimum amount of effort I was able to get 6.1.3 installed in an Ubuntu 22.04 install in WSL2 and it ran smoothly. Did the below full size (832x1248) SDXL image in approx 64 seconds.
kiwis_in_the_night.png

Formerly that would have been over 7 minutes on my Windows install. I haven't had a chance to reboot into Linux and try the same image on raw Ubuntu with same hardware for comparison. My feel is that it will be slightly shorter but not by much. IIRC, I could get a full SDXL image no scaling in slightly under a minute. I'll do a proper comparison when convenient but the take-away is that Stable Diffusion using AMD on Windows has just had a major leap forward and become much more viable. @Susanna should be happy as she can now have everything in containers as I know she loves! That is if she hasn't deleted Windows entirely these days.

That said, there's something not quite adding up here. I ran a non-WSL version for comparison and I was expecting it to run at the same slow speed it used to but it was no fast as well. My only theory is that in updating the AMD drivers which I did for the WSL2 version, it brought major improvements to my pure Windows version too. I'm a little confused because I don't know why there aren't huge differences due to ROCm. Or perhaps I'm simply going wildly wrong and it's just using DirectML for both and there's been some huge leap forward since I last tried it. I'll figure out what's going on.

Still, good to know that AMD is now viable on Windows for this.
 
  • Informative
  • Feels
  • Like
Reactions: The Mass Shooter Ron Soye, indomitable snowman, Baraadmirer and 1 other person
Catgirl Tyranid said:
View attachment 6110119
3 attempts, all failures. I literally just want this avatar, but with snow falling.
Click to expand...
It's a bitmap style image. Have you considered just opening it in a graphics program, duplicating it to 15 layers and drawing some little white pixel blocks on it? Heck, you could even have them all as one transparant image and paste it on to each duplicate layer in turn but 30 pixels lower, and then you have lots falling with minimal work.

Crop to original size and export the layers as GIF. Thoughts?
 
Overly Serious said:
It's a bitmap style image. Have you considered just opening it in a graphics program, duplicating it to 15 layers and drawing some little white pixel blocks on it? Heck, you could even have them all as one transparant image and paste it on to each duplicate layer in turn but 30 pixels lower, and then you have lots falling with minimal work.

Crop to original size and export the layers as GIF. Thoughts?
Click to expand...
I hear you speaking logic, but it conflicts with my desire to repeatedly bang my head into a concrete wall like a retard.

But yeah, I'll probably do this when I have time. I just wanted to see how hard it was to get simple animation effects out of it.

Unrelated:
 
  • Like
  • Feels
Reactions: Jimjamflimflam, The Mass Shooter Ron Soye, Baraadmirer and 1 other person
Overly Serious said:
@Susanna should be happy as she can now have everything in containers as I know she loves!
Click to expand...
I already have my stable diffusion setup containerised though? And I only use windows for Solidworks and a handful of games that are just too much hassle to set up on Linux.
I’m happy for AMD windows users, of course, but this update isn’t a big deal to me. Especially since I got a 4090 a while back when I got into chatbots, the 6900XT has been gathering dust since.

I’m trying to use python to tie my chat bots together with a voice recognition software and a text-to-speech library so I’ll have a LLM I can simply talk to. I still spin up SD every now and then, and I might try to set up my chatbots to prompt it, Bing style, but for the most part I’ve moved on.
 
Susanna said:
I already have my stable diffusion setup containerised though? And I only use windows for Solidworks and a handful of games that are just too much hassle to set up on Linux.
Click to expand...
I was semi-joking but making the point that your desired approach no longer keeps you from doing it on Windows.

That said, I found a moment to do some performance tests and I clearly misremembered my response times on Linux or else something has changed to improve them. An image that takes 60 seconds on Windows takes me, with exact same parameters, around 20 seconds on the same hardware.

So whilst AMD have added extensions which now allow one to use ROCm code from within WSL (and presumably therefore a container), it's clearly not running in the same way. So... whilst my posts weren't a false alarm - this is a very positive sign, it's not freeing me from dual boot just yet.

Susanna said:
I’m happy for AMD windows users, of course, but this update isn’t a big deal to me. Especially since I got a 4090 a while back when I got into chatbots, the 6900XT has been gathering dust since.

I’m trying to use python to tie my chat bots together with a voice recognition software and a text-to-speech library so I’ll have a LLM I can simply talk to. I still spin up SD every now and then, and I might try to set up my chatbots to prompt it, Bing style, but for the most part I’ve moved on.
Click to expand...
I can imagine you very happy as the last human being on Earth, saying "Mir at last!" as you fill up your generator with a lifetime's supply of Russian gas and build yourself a family of amiable robots.

XL xQgg?QcQCaTYDMjqoDnYpG said:
If you want to do stress tests, try my avatar
lol
Click to expand...
I would but it's probably a Magic Eye picture with a hidden cock in it. Would lose me a luma account. ;)
 
  • Disagree
  • Feels
Reactions: XL xQgg?QcQCaTYDMjqoDnYpG and Susanna
One of my pet peeves about AI art are people so lazy and incompetent they can't fix the fucked up hands and facial features and other oddities and just post the raw image as a finished product. I think the ideal balance at least right now is someone who doesn't deny that AI has a place but when needed/wanted still has the raw artistic skills to go in and work with the picture themselves. A combination of new and old talents so to speak.

Some AI users on the other hand seem completely stumped and have to spend hours regenerating an image over and over to get something right that could easily be fixed in photoshop with traditional skills in a few minutes. That is if they don't take the lazy route and just post the unfinished image.

While I don't agree with the antiAI camp that anyone who uses AI is not an artist I do agree if you're utterly overdependent on it that you can't function at all outside of it you probably arent an artist.
 
  • Like
  • Agree
  • Feels
Reactions: AnimeSucksCopeAndSneed, Anti Alien Alarm, Catgirls are Love and 3 others
So on the political and legal side of things, SD3 is turning into a real debacle. Civitai has had their lawyers go over the SD3 licence and Civitai is sticking with its wholesale ban on SD3, SD3 derivative works and I think even SD3 images.


(from the article)
1719160796245.png

Between the technical issues on one side and the legal scare going round on the other, this release is looking really bad for Stable AI. Frankly, I think they're going to have to redo their licencing terms though I don't really know what internal constraints and other opportunities they might have. I mean, Civitai is a site with a userbase that kind of leans in a certain direction (i.e. I keep my filters ON on on that site! :/ ). But it's also the largest community and place for customisation and fine tunes of the stable diffusion models. Not having that level of community engagement, especially when SD3 has such technical issues requiring fixes, can't be a good thing for its success.

I've seen more talk about Pixart and Hunyuan and Lumina models in the past week than in the previous year. People are not impressed. Maybe it'll blow over. I've said it elsewhere, the 8B version of SD3 seems good and the design of SD3 is interesting. But right now I would not want to be in charge at Stable AI.
 
  • Informative
  • DRINK!
  • Agree
Reactions: indigoisviolet, woowie queen, anustart76 and 3 others
You must log in or register to reply here.
Back