Stable Diffusion is one the newest AI image generators and is completely open-source. Learn how to use it for image creation, upscaling, and more with our detailed guide.
https://www.techspot.com/guides/2590-install-stable-diffusion/
How to Run Stable Diffusion on Your PC to Generate AI Images
- Thread starter neeyik
- Start date
Hodor
Posts: 765 +512
I tried stable diffusion the other day (hosted on some site). My prompt was "kangaroo driving a sports car". I'd like to paste the resulting image here, but there's no option to upload a local image. Anyway...... it's two stylized kangaroos in a sports car, and the second kangaroo seems to be playing with a certain XXX body part of the first kangaroo.
So, either the admins trained the AI using PornTube, or the AI has dirty imagination. Whatever the reason, I like it.
So, either the admins trained the AI using PornTube, or the AI has dirty imagination. Whatever the reason, I like it.
UdyrL
Posts: 147 +234
"Intelligence has been defined in many ways: the capacity for abstraction, logic, understanding, self-awareness, learning, emotional knowledge, reasoning, planning, creativity, critical thinking, and problem-solving. More generally, it can be described as the ability to perceive or infer information, and to retain it as knowledge to be applied towards adaptive behaviors within an environment or context."
To some extent, yes... this is intelligence, limited when compared to humans way of processing, but intelligence nonetheless.
The Git repo has instructions for AMD cards, and indeed, one of the arguments that you pass with those instructions is "--no-half". The instructions are for the Linux operating system, not Windows. It does link to a Windows instruction manual, but it says that it is untested. Neither of these are particularly surprising, a majority of AI work is done on Linux and on Nvidia GPUs, particularly CUDA. I'm surprised there's even a manual for AMD cards, which is a big plus in my opinion. Anyways, here's the link: https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-AMD-GPUs
I've been using SD WebUI for months, running local on my PC.
It's great to also add some of the many scripts and extensions that have been developed for it.
Another great thing is to use the other models derived from SD, refined on other image datasets with other styles. for example StudioGibliSD, WaifuDifusion, TrinArt, ArcaneDifusion, several specific ones to create pixel art, and others.
also the integration with other applications such as Photoshop or Krita
In general, if you want to take advantage, it is necessary to study the interface and its many options, the generation techniques, the "prompt engineering", ect
I have even collaborated on github, reporting errors or suggestions that have been taken into account.
I've even been planning to redo all the art from a little point&click adventure game I did over 15 years ago.
It's great to also add some of the many scripts and extensions that have been developed for it.
Another great thing is to use the other models derived from SD, refined on other image datasets with other styles. for example StudioGibliSD, WaifuDifusion, TrinArt, ArcaneDifusion, several specific ones to create pixel art, and others.
also the integration with other applications such as Photoshop or Krita
In general, if you want to take advantage, it is necessary to study the interface and its many options, the generation techniques, the "prompt engineering", ect
I have even collaborated on github, reporting errors or suggestions that have been taken into account.
I've even been planning to redo all the art from a little point&click adventure game I did over 15 years ago.
there are some variants to be able to use AMD, in windows and linux, either using ROCm, SHARK (for RDNA) or DirectML. but the work is without user interface, it is modifying the prompts directly in the Python code
there are some variants to be able to use AMD, in windows and linux, either using ROCm, SHARK (for RDNA) or DirectML. but the work is without a user interface, it is modifying the prompts directly in the Python code.
or you can run it on the CPU, much slower of course, just need to specify one parameter in the startup file so PyTorch won't check if there is a CUDA device : --skip-torch-cuda-test , also use --no-half
AxelAminoff
Posts: 28 +24
That worked, indeed!
Thanks. I was só disappointed yesterday, with doing everything and ending up with that "Half" argument. Didn't know it was this simple, though. So, thanks for completing that last 1% to get it up & running!
I also tried this guide, but it didn't work for me: https://pythoninoffice.com/run-stable-diffusion-amd-gpu-on-windows/
I couldn't get the huggingface-cli login to work ("Invalid Token")
I can't wait to see guides coming uit to use my AMD GPU, though.
Now it's CPU-bound.
(I am seriously happy with my -very old- R9 295X2, with 2 GPU's and 2x4GB of GDDR5)
Tinckerbel
Posts: 29 +46
For those that like an easier to understand GUI and don't want to fiddle around with Python, GitHub and command line inputs and still want to achieve good results on Windows, NMKD Stable Diffusion GUI is a must have:
https://nmkd.itch.io/t2i-gui
The results I got with a custom ckpt model are stunning.
https://nmkd.itch.io/t2i-gui
The results I got with a custom ckpt model are stunning.
WhoCaresAnymore
Posts: 775 +1,223
I'm utterly against the AI takeover of Art. Is there no place for human creativity anymore? We designed the freaking AI that's taking over.
Everything seems so prophetic.
Everything seems so prophetic.
Thanks for the suggestion, we're giving this a try and possibly adding it to the guide soon.
Hodor
Posts: 765 +512
why don't you give it a try? probably it can!
alexnode
Posts: 149 +56
Yes , I am quite curious to see how this will affect society as a whole. It is still not very consistent ... SD does it's own thing.if you really want romething in a specific room you might need three pages of prompts. I find it really dubious that you could input the name of an artist so you could imitate their style. If you were doing that in music you would have 20 lawyers on your doorstep. But art and photography are at a weak state. Next step is to replace software developers...
Tams80
Posts: 218 +168
ChatGPT is scarily close. You can get it to churn out code for a problem very readily. Now, the code often has some issues, but nothing that often a quick clean-up can't fix by someone with only limited coding knowledge.
m4a4
Posts: 3,809 +5,446
It's not taking over anything meaningful. And it still takes human input (creativity) to generate anything worthwhile.....
m4a4
Posts: 3,809 +5,446
It's not. It can pump out simple code with accuracy, but anything too complex and it'll confidently get it wrong.
And no, it still needs a competent programmer to fix the bigger issues and make sure it's up to par...
I think it will get there, it's not a matter of if but when. That aside, Radeon cards don't see a good support in stable diffusion under windows. It works with openML. But under linux this webui runs much faster (with rocm), requires some patience for installation though. AMD should take AI more seriously and implement newer tech to better accelerate this kind of workloads or Nvidia will pull further ahead.
I should have made it clearer in the article that the LAION-5B dataset isn't really comprised of images -- it's a list of the URLs for the original images and the alternate text that's describing them. The project looked at each picture and its description, calculated the mathematical link between them, and then used that data for the training. Essentially, it's no different to an artist looking at a painting of something and using it for inspiration to create their own piece of work -- ethically, one could argue that said artist should thank the original creator for the inspiration, but there's no legal precedence to do so.
Similar threads
