Install Stable Diffusion on windows in one click! AMD GPU's fully supported!

Want to run Stable diffusion on windows with an AMD gpu? Install and run Shark from Nod.ai in one click. Simplest way to get stable diffusion up and running even on AMD.
Page to download installer:
github.com/nod...
Direct link to installer:
github.com/nod...

[@EvyDevy]

Thank you so much!

[@Pelorso]

thanks for useful video, finally SD works with my 6600

[@FE-Engineer]

Whahoo! Glad it helped!

[@Torva01]

How many it/s? Mine runs with 2,4 it/s and only 512x512 :(

[@akshay12341]

Thanks for this. Explained everything very perfectly.

[@FE-Engineer]

Thank you for watching! I hope this helped!

[@parryhotter3138]

looking forward to the comfyui vid. installed it but have not tried it out so far. would be awesome to know whats possiible and what not with amd. i really like that u are going a bit more into specs and troubleshooting... most of the other creators are just telling you how awesome it is and what u can do with it... but not that its possibly not even runs on your system. so thank u, finally inpainting... ;)

[@FE-Engineer]

Your username on here makes me laugh every time I see it. 😂. I promise. I’m getting to comfy ui. It’s such an unusual program. But I am getting there.

[@ai.d__g]

For AMD, Stable Diffusion Direct ML works really well. Also allows for ControlNet which is really helpful for tiled upscaling.

[@FE-Engineer]

Does controlnet work on the direct ML version? I have it installed on my version in Linux I was not aware that it also worked in directml version since most things honestly do not work in that version.

[@ai.d__g]

@@FE-Engineer yes control networks on Direct ML .

[@FE-Engineer]

Sigh. I’ve installed controlnet. Gotten the controlnet files. And tried using it several times. For some reason numerous people tell me it works. For me sadly it has never worked. :-/

[@ai.d__g]

@@FE-Engineer you've installed the controlnet UI ?

[@Knightedskull]

How did you get your SD to work for your card? It always switched to CPU for me.

[@Krautrocker]

Appreciate the videos so far, helped me very much setting SD up on my new AMD rig! Couple of questions though: Will you make a guide videos on weighting/emphazising prompts? Is that even possible with directml? Do "weighted" Loras even aplly with AMD directml? I just started with ai art in general, so my Nvidia using friends provided me with models and Loras. The models i could optimize easily enough with your video about them, but i cant seem to get any Lora applied and changing the weight numbers doesnt seem to do anything.... Sorry if im a little bit all over the place and asking too much in one comment

[@FE-Engineer]

Sure thing glad it has helped! Emphasizing and weighting prompts should work without any issues. I’ll double check though. As far as Lora’s or weighted Lora’s I’m not sure about. Unfortunately the automatic1111 directml version a fair number of things do not actually work. :-/

[@thelaughingmanofficial]

@@FE-Engineer Not sure if it works with Shark but I know in Vladmandic if you put a prompt in ALL CAPS it put more weight towards that prompt, at least that is one way to do it, there are other more elaborate ways to do it as well. But the simplest is ALL CAPS.

[@nenadm5747]

@@FE-Engineer It works...

[@FE-Engineer]

Generally when I want to weight a specific prompt. For instance full moon. I would do it like this (full moon:1.1) and go up higher 1.2, 1.3 etc to add even more weight. Most things even 1.2 or 1.3 is adding a lot of weight to that specific piece of the prompt. Besides that. You can also manipulate the CFG scale as well to be less creative and more strictly adhere to the prompts. In automatic. If you select a group of text and press control and up it should automatically start applying higher weights to it.

[@youssefLKHL]

I really really appreciate the well explained method to get it running, i have 1 question though: is there a way to run comfyui on windows using an amd card I'm running a rx 580

[@FE-Engineer]

Ohh. Comfyui. I am still trying to get a video on this. At this moment. I don’t believe there is any reasonable way to get it onto windows with an AMD card. When ROCm is fully out on windows with MIOpen and MiGraph (soon?). That will very likely change things and make it possible. I’m currently testing and working with ROCm 6. If you want it right now, your easiest and best bet is ROCm on Linux.

[@youssefLKHL]

@@FE-Engineer is there any way to make comfyui running on linux using Rocm 4.0, as I believe this is the latest supported version for my rx 580

[@hhssw_4057]

its alive!! Thanks a lot! Worked with 7900xtx. Speed of generation is really fast

[@FE-Engineer]

Glad it worked without any issues. :) you are welcome!

[@hhssw_4057]

@@FE-Engineer By the way, I was able to run chat on the llama2 7b model in the sharkai. The generation speed was quite high, although there are no settings in the software (that’s why I asked to look at oobabooga)

[@martinhalaj9604]

does not work for me. It says "could not compile Unet"

[@SorenzoAzhathot]

When i doble click the exe nothing happens

[@FE-Engineer]

Umm. Sorry to hear that.

[@SorenzoAzhathot]

@@FE-Engineer can I do anything to fix that?

[@SorenzoAzhathot]

@@FE-Engineer can I do anything to fix that?

[@FE-Engineer]

Comfyui doesn’t have an exe file. I’m not really sure what you clicked on?

[@SorenzoAzhathot]

​@@FE-Engineer 2:38 ?

[@4MERSAT]

test support for SDXL models is available in the latest shark beta releases. But I can't say anything about it, because I haven't tested it myself.

[@FE-Engineer]

Sweet. That will be nice once that functionality is available.

[@paul2744]

Help, the first time all worked, but i close the application for install model, reopen it and don't open, i started to uninstall/install a second one, it don't want to open, when i double click on it, the terminal apear and deseaper ... someone can help me ?

[@PatientXero607]

Holly crap. 10 seconds for 512x512 at 50 steps on my RX 6800 XT. Incredible! 3.39 it/s per second with Dreamshaper v8. The only problem that I with multiple types of models is that it will not make an image with 768x768. The images are garbled. 512x512 works fine. Edit: OK, 512x512 for SD 1.5 and 768x768 for SDXL. XL is 1 it/s with 50 steps.

[@FE-Engineer]

Sdxl is significantly slower. Plus larger images is significantly more pixels to render. Resulting in even slower. So I am not overly surprised. I as well see a 60%+ performance drop with sdxl.

[@RaZziaN1]

for me it takes 200 seconds.. for 384x384 (smalles possible) image on 6750xt.. on most models - huggingface/civitai ones

[@RyanAdkins-q9l]

thanks for useful video Awsome！！

[@jfrost5096]

when reopening the Shark Studio UI, do you go into the Shark Folder you created and downloaded everything to and then open the "nodai_shark_studio_20231009_984" application? I can't get new models to appear after dropping the file into the models folder

[@FE-Engineer]

I have not messed with shark in a while. Is there a reload button next to models?

[@larion2336]

I hate the constant downloading and duplications of files. When I tried it I wound up with 40+gb of files from just changing a few parameters and trying 2-3 models. Think I will just stick to doing it on Linux via dualboot, I'm already running LLM's that way via text-gen webui. Running a1111 on linux works without any of the duplicating crap, everything works. It is technical to set up though.

[@FE-Engineer]

💯, having it in windows is nice. It does work. But too many things do not work, the crazy amount of duplication of models. I prefer my Linux version that does everything is maybe 5-10% slower but works out if the box for everything once setup properly.

[@nenadm5747]

I aready use Direct ML version for a few months and everything works, Lora's, Controlnet, Roop, Adetailer, Paint, all the models...the only thing that bothers me are the resolutions, the max I can get is 540x960 and less if I use the option --no half, because the paint feature works only with that option... Btw, my Gpu is RX 6800M, it's a good card but because of lack of Amd support, it's limited. Is this version any better? Faster maybe? Regards

[@FE-Engineer]

No. Definitely not. And right now almost everything is broken. So if you have controlnet and stuff working don’t update it.

[@FE-Engineer]

Running in ROCm is of course the best way to go. As that lets you do everything.

[@nenadm5747]

​@@FE-EngineerBut that's possible only on Linux...for now at least, right?

[@FE-Engineer]

Yes. For now. ROCm is Linux only. Judging from GitHub progress it does look like it is getting close though.

[@nenadm5747]

@@FE-Engineer Good news 👍 Meanwhile, I can install Linux on my ext hard drive, to try ROCm version...I hope it will not be slow, it isn't SSD, but it's 3.2 gen 1...

[@Thedeepseanomad]

Do you think they will add stuff like dreambooth and such? What is the best Radeon solution that includes dreambooth? Something you have to run on Linux?

[@FE-Engineer]

ROCm on Linux currently is by far the best. But ROCm 6 is coming out this month. And it will hopefully be available on windows. If not when it comes out. It should support windows soon. Whenever that happens it will make AMD much much better and easier on windows.

[@leucome]

Dreambooth run on Linux with ROCm. But last time I used it the dev made mistake that disabled previews on AMD and Apple. So I had to stop the training to generate a preview manually. The training itself was working as expected. Maybe it is fixed by now.

[@ToddGoo]

Finally, my 7800xt GPU runs stable diffusion . Thank you very much. How can I use Controlet under shark nod ai? thank you for your help.

[@FE-Engineer]

Shark is still working on integrating controlnet to my knowledge.

[@user-vg6rp1wn8l]

Did you use Vulkan? I haven't seen any software using Vulkan for GPGPU since Waifu2x.

[@FE-Engineer]

I did before. It worked fine but was super slow. Fast to render. Slow to recompile shaders for numerous types of configuration changes such as changing the output size.

[@ViralWatchMedia]

Just a fYI this doesnt work anymore, the installer never gets past the first part and a white screen pops up in the shark UI and nothing happens.

[@FE-Engineer]

That seems hard to believe. AMD bought out nod.ai which is the company that built and maintains shark. I have a hard time imagining that they would let their product not just work for AMD. But I’ll take a look. I appreciate the callout.

[@FE-Engineer]

Are you sure you have all the required dependencies installed correctly?

[@newgamer1359]

How can I use shark nod ai from my local network? I'm new to all of these things and I want to access this with my phone from my pc

[@dropfry]

3 weeks after this video, How badly did I screw myself by buying a 7900XTX? Or have there been any improvements made since this video?

[@FE-Engineer]

oof...that is a loaded question. My personal take -- Nvidia cards are just too expensive with not enough hardware to backup their prices, namely VRAM is lacking AMD, significantly more hardware for your dollar but software is lacking I have videos for using Automatic1111 DirectML, ComfyUI, and Shark from nod.ai in windows Each of these pieces of software have some drawbacks for AMD users in windows. ROCm on linux -- is awesome, and you can basically do everything that Nvidia users can do without really much of any limitations. So if you are game for a dual-boot system with linux, you are golden, otherwise you will unfortunately have to deal with some limitations in windows until ROCm for windows finally comes out (hopefully before too long, most of the work is done it seems).

[@PanKrewetka]

Is there any option to use ComfyUI with 6600 XT? No matter, is it Windows or Linux.

[@FE-Engineer]

Yes, I will have a video for installing and running ComfyUI here hopefully this week. Comfy is definitely the most complex stable diffusion to use, so I saved that one for last. Keep an eye out though, I will try to get that one done next!

[@PanKrewetka]

@@FE-Engineer thank you so much, I'm trying since morning with installing Linux, but drivers go messy with ROCm and I'm total newbie. Can't wait to see that video, thank for all your work!

[@FE-Engineer]

Sure thing! Glad you are finding it helpful, thank you so much for the kind words!

[@emanuelager5627]

once everything has installed, the web.ui started. But when i tryied to start the exe again the Web.ui and the cmd did not open any idea what could be the problem

[@FE-Engineer]

Reboot the computer and try again. It is probably already running and you are simply not seeing it in the taskbar. The icon in the taskbar is not something that is easy to recognize.

[@wassaysarwar]

I want to use Fooocus on AMD. Do you have a tutorial for that?

[@FE-Engineer]

I do not, however after looking closer at it, I think this will be an interesting video especially for windows AMD users. Let me do some research and get it figured out so that I can make a reasonable video on it.

[@fhuckgoogle4171]

I'm stuck on "Compiling Vulkan shaders. This may take a few minutes." I'm using a 5700XT and it's been 15 minutes. I've tried reinstalling and restarting my computer but it just won't move past this point.

[@FE-Engineer]

:-/ that’s aggravating. I’m sorry to hear that. I do find the compiling shaders to take a very long time even on a 7900xtx. :-/

[@fhuckgoogle4171]

@@FE-Engineer Thanks for your response. Do you think it might take an hour or more? I’ve tried at least 10 times but always the same result. I even left the house for half an hour and when I came back it was still compiling. I’ve given up at this point and am going to try another Stable Diffusion alternative. Although every single one I’ve tried is either too complicated for me to work out or it just won’t work for some reason. I don’t really want to waste a couple of hundred dollars on an nvidia gpu that will only be slightly more powerful than my current.

[@hhssw_4057]

Could you record a video on how to install and run models for text generation on webui oobabooga via windows 11 on amd gpu?

[@FE-Engineer]

I’ll look at it again. I’m pretty sure it requires ROCm which is not available in windows yet.

[@FE-Engineer]

Shark runs text generation. It seemed off to me though. It didn’t seem to be working appropriately. It was talking entirely in slang. You could try running shark if you want. I just put up a video about running shark. It’s in windows. Literally one click install. Dead simple. I just found the chatbot to be really odd and it seemed like it was not working appropriately. It also was slow. In Linux I can get 50+ it/s. On windows it was getting 1.5 it/s. Hence why I recommended folks not to really use it.

[@FizzyPinepple]

I get a SystemError: Coul not compile Unet both with the .exe on windows as well as with the dev build on arch :( And the directml fork of AUTIMATIC is so slow i might as well just use the original and ran it on my cpu lol

[@FE-Engineer]

Have you used zluda? If the directML fork is as slow as using cpu it sounds like something is wrong

[@cutza7]

Why did you say rocm doesn't work? Is that function coming up at some point?

[@FE-Engineer]

I said it because ROCm is not available in windows…it is not a function. It is an enormous library of software that basically tells the graphics card how to handle AI workloads.

[@molodoymolod]

tried shark before and tried it now again. it still doesn't work

[@FE-Engineer]

strange, AMD bought Nod.ai (the company behind shark). So I am pretty confident that making sure AMD cards work with Shark is not optional for the company. My guess is you may have some very odd things going on with software in windows, multiple python versions installed, or something along those lines causing some problems. For most people it seems to "just work". Sorry to hear it did not work for you though. :-/

[@PsychoBenches]

ay we got the same GPU. Thanks for this bro edit: yeah same cpu too lmao

[@FE-Engineer]

Happy to help! Heh I love both my cpu and GPU. I needed a new computer and this was the best performance I could get without spending like $4000.

[@PsychoBenches]

@@FE-Engineer Which AIB did you get for your 7900 xtx?

[@doug_durray]

Thanks so much for this, only have a 6800xt but seems to work quite well! But anyone know how to get image 2 image to work in this. It seems to completely ignore the image I put in rather then alter it?

[@FE-Engineer]

So I think you might be using it a little incorrectly. If you want to edit part of an image. Inpainting works quite well. It keeps most of the image and repaints a small piece according to your prompt. Image to image in general is hard to use but refers to using the first image as a bit of a basis but then draws a new image according to your prompt but adding your initial image in the calculation.

[@FE-Engineer]

So in general. If you want a whole new image but want one that is somewhat similar to the original. Inpainting is if you want the original but want to change or refine a specific part (like an ugly face or distorted hands in particular)

[@Aquilenne]

In addition to the inpainting thing to selectively target things, I'd like to add that lower denoising also helps to avoid having the program generate something wildly different.

[@milosdraskovic4944]

try to decrease denoising strenght

[@Janekdraws]

Super helpful as always, but why does this have to be what I am watching after spending half a day getting Linux and all the drivers on my tiny ASUS Rog Ally :')

[@FE-Engineer]

Oh Linux is the way to go for sure. This is really for folks who want stable diffusion with absolutely no hassle. Windows only. No command line. Etc. it’s great for what it is but it definitely doesn’t really compare to running rocm with controlnet and dreambooth and all the bells and whistles. Plus it is absurdly annoying and time consuming to just change models.

[@FE-Engineer]

Next video is probably comfy ui. That will be fun in Linux :)

[@griggpev]

hmmm only seem to get around 5.62 it/s with a rx 7800xt. thoughts?

[@FE-Engineer]

What size image are you generating?

[@FE-Engineer]

That seems low to me though. I would think you would be getting probably around 10 or so at 512x512

[@griggpev]

@@FE-Engineer exactly as you say, 512x512. I thought I’d get more too :/ but it dropped to 3.4 it/s with further testing… strange

[@Hecbertgg]

i have 2 options: 7900XT vulkan 7900XT rocm which one is better?

[@FE-Engineer]

Rocm although I didn’t know it worked with Rocm.

[@FE-Engineer]

But if you can do it with rocm. 100%. Rocm.

[@WillPl4Y]

Is it only generate image? Or video too?

[@FE-Engineer]

Stable diffusion generates images. There are a few extensions that do “videos” but this is not much more than taking an image and moving the background and the foreground slightly differently and getting like a 3-5 second video. So yes it can do video with some extensions but personally. I barely would qualify that as “video”.

[@ushiro2031]

the lastest amd driver is 21.1.1, it will work?

[@FE-Engineer]

I would assume so. It should. Also new driver came out today I think.

[@DJVibeDubstep]

Does this still work? I heard of recent updates breaking things. I have a RX 5700 XT and I'm forced to use my CPU right now to use SD and its painfully slow as fuck lol.

[@FE-Engineer]

As far as I know yes. But if you want it to really run fast and have everything work. You really should run rOCm on Linux.

[@DJVibeDubstep]

@@FE-Engineer thank you for the fast reply! I'm still somewhat of a noob with this AI stuff. I have Windows, how can I run it on Linux? I'm not sure what rOCm is either I'll have to look into that!

[@tidedetergentyummy]

Is this locally run? I want to run this locally. I'm new to this.

[@banned-user]

Yes it is run locally

[@tidedetergentyummy]

To create a public link, set `share=True` in `launch()`. Idk what this is.

[@banned-user]

@@tidedetergentyummy you can use that and it will create a gradio (I think it's called that) link which you can enter on another device and use SD from anywhere. Be aware that it is a slight security vulnerability but you can set a password on it

[@FE-Engineer]

Generally speaking I don’t really get into running stuff from external services or cloud things much or often. 99% of my channel is focused on running things locally.

[@Laiirton]

omgggggggggg

[@FE-Engineer]

:) thanks for watching!