Stable Diffusion Works on AMD GPUs - Step by Step Windows Install Guide (EASY!) 

Thank you, glad you are up and running!

You're welcome!

Hi, I tested on SDXL. I was getting a very black image on Stable Diffusion 1.5.1 using SDXL Base 1.0. I did a git pull and upgraded SD to 1.5.2 from the repo. It is better but I still have my backgrounds turning out black in most pictures on 512x512. I don't have enough video ram to allow 1024x1024 sizes and I think that might be my problem. I will play with this a bit. I also followed these instructions for the SDXL model. stable-diffusion-art.com/sdxl-model/#Download_SDXL_10_model

You're welcome!

Hi, Your card has 12GB of RAM. You should be able to run without any of those settings. I see some use only --no-half-vae. The --medvram is for 6 to 8GB and not needed for cards with lots of VRAM. Check out this post for some discussion on RX6/7xxx settings. github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/8626 Let me know if you come right.

@@Celenwen Have you found a solution?

Hi, I am running it with the latest standard drivers from AMD on an older RX card. No special drivers. It works well. Try and installing as per instructions in the video. You need Git to clone the repo I use in the video.

yes thats exactly what i am doing, pictures to put in my audio lessons for my students, at the end the pictures don't have to be super great 2k or 4k but just enough to convey my lessons so this should work out for me.

Thanks!

Hi, Thanks for reaching out. There seems to be a bug on the Automatic1111 AMD repo. Can you try this video, its a workaround that should work until they sort out repo. Let me know if it helped. ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-TJ98hAIN5io.html

Hi, Thank for reaching out. Are you getting any errors when following the video? What OS are you running on and let me know what Arguments you are using to launch SD?

Hi, Thanks for reaching out. There seems to be a bug on the Automatic1111 AMD repo. Can you try this video, its a workaround that should work until they sort out repo. Let me know if it helped. ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-TJ98hAIN5io.html

Running stock standard, will check it out. Thanks!

Hi, Thanks for reaching out. The only setting I use for my card is the --medvram setting. I do think you need the correct version of Python here. Let me know if you that sorts it out..

@@stuffaboutstuff4045 Hi bro. I'll tell you: change the python version to 3.10.6 but the gpu still doesn't work. I have tried with a released version of automatic1111 and neither. Now I have options like trying the focus interface that comes with the XL model, or installing ubuntu which, according to what I've read, does have the ability to work with AMD GPUs. what would you recommend? I'm also a linux user so it wouldn't be a problem for me. I just would spend more time doing the installation. Thank you greetings from Venezuela.

Hi, you can try this fix posted by @karikaturdigital6123 on the SD Fix Torch Video. Check the video on the channel. ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-TJ98hAIN5io.html Fix for who has problem with No module named 'keras.__internal__' cmd on sd web ui directories venv\Scripts\activate pip install onnxruntime-directml pip install torch-directml pip install keras pip install tensorflow

Hi, Currently looking into into it. I will let you know when I have something ready.

@@stuffaboutstuff4045 Great!

Hi You must use the exact version of Python I use in the video. Also make sure you execute all the command from "admin mode" command prompt. Let me know if you came right...

Thanks for the feedback!

hey do you still use stable diffusion? i have same GPU. which model should i install? Shark or Dreamshaper?

I am using the Dreamshaper model.

Hi, Thanks for reaching out, are you getting this issue on Windows 11?

@@stuffaboutstuff4045 Same error on windows 10 pro here. (thanks!) radeon 6600xt, new install of windows 22h2

@@matthewfuller9760 this is also something of note it is fast but using my cpu, im on a 5800X, very strange as its working and generally only takes a couple minutes im not sure if this is cause for concern atm. I am also on win 11 22h2

@@BulbOfSalad I don't think it requires win 11 though?

@@matthewfuller9760ofc not.

are you trying to (run it with administrator) settings??

did you cd\ to get into the root of c:\? Command prompt will open in Windows\system32 then just type cd\, let me know if this resolves the issue?

Hi, can you check if this video applies? ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-TJ98hAIN5io.html Otherwise share your command line arguments used for the GPU. Let me know...

Hi, Mine only worked in admin command prompt. If you launch like I do in the video you should be good to go. When it runs the first time it will install and set the environment up. Watch the screen for any errors. Let me know if you get it up and running...

Hi, I remember having similar issues, I ultimately resolved these issues and the two main components here were make sure you use the exact version of python recommended and run everything in administrative command prompt window. Let me know if it works. After ensuring this it was a pretty smooth install and run.

@@stuffaboutstuff4045 Professor, thank you so much. Thanks to you, I solved it. No matter how many times I say thank you, it's not enough.

Pleasure!

Thanks!

Hi, What settings are you using in your config file. It must be issue in the config file telling it to use the CPU. Let me know if you come right.

@@stuffaboutstuff4045 Thanks for replying. Can you please specify the address of the config file? There are some files on my stable-diffusion-webui that have "config" in their name, such as configs or configs.json.

Hi, honestly I did not test performance on ROCM. I tested on default Windows driver and on the latest AMD driver. I could not pick up a difference between those two. Should you have loaded and it does boost performance please let me know. Granted I am running RX580

@@stuffaboutstuff4045 i installed rocm, but i turned off ai acceleration because i dont want SD to break.

Hi, thanks for reaching out. I did in the second video. ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-TJ98hAIN5io.html I will build SD again and maybe just make an updated video with consolidated current steps..

@@stuffaboutstuff4045 Thank you very much.

This was run on 8GB RX 580 Radeon. My pressure is purely on the GPU. Not super fast but not painfully long waiting times as well. Glad to hear you are up and running.

@@stuffaboutstuff4045 thanks for the video!

i have a 4gb RX 570

Here's the log: Traceback (most recent call last): File "I:\Modelo_IA\test 2\stable-diffusion-webui-directml\launch.py", line 48, in main() File "I:\Modelo_IA\test 2\stable-diffusion-webui-directml\launch.py", line 39, in main prepare_environment() File "I:\Modelo_IA\test 2\stable-diffusion-webui-directml\modules\launch_utils.py", line 560, in prepare_environment raise RuntimeError( RuntimeError: Torch is not able to use GPU; add --skip-torch-cuda-test to COMMANDLINE_ARGS variable to disable this check

Hi, Just want to check, you did follow all the steps and clone the AMD SD installer as in the video. You have to use the exact version of Python 3.10.6 (Add Python to Path when you install it). Run all the commands in "Admin" mode command prompt. I have tested this with both the Windows driver and AMD drivers and it worked on both using RX card. Also on a 4GB RX 570 if you follow all the steps in the video you should only need to use the following command line arguments. COMMANDLINE_ARGS=--opt-sub-quad-attention --lowvram --disable-nan-check. Let me know if you come right...

@@stuffaboutstuff4045 Yes, I followed all the steps. In the step that opens webui-user, this error appears as the last message. If I don't add the command --skip-torch-cuda-test it doesn't even open the browser

I had python 3.11 and I downgraded it, could it be that I did something wrong?

Hi, bit of late reply, can you clarify the question..

@@stuffaboutstuff4045 he asked about DARK THEME

Hi, You have a lot of video memory. You should be able to run SD without any special video RAM settings and using the defaults for the webui-user.bat file. If you check task manager, is it using CPU or GPU when creating? Is it slow on all models you try? Let me know if you come right...

I have the exact same issue bro

I Think SamaSumi and I have similar issues, so we would have to apply "--no-half" command to the bat file, because if not, it would show errors that some files not found in file 'Half". However, if we do apply the "no half" command, it will run without using the graphics card, it's gonna be very slow@@stuffaboutstuff4045

Hi, RX 580...

Hi, Thanks for reaching out. I know I has some issues with the build that was solved that by running everything in the admin command prompt as in the video. It is also important to ensure you use the correct version of Python. Did you further follow each step in the video? Let me know if you come right with this problem.

@@stuffaboutstuff4045 What is the correct version of python? I'm using version 3.10.6

Hi, Thanks for reaching out, are you getting this issue on Windows 11?

yes sir@@stuffaboutstuff4045

Hi, Thanks for reaching out. There seems to be a bug on the Automatic1111 AMD repo. Can you try this video, its a workaround that should work until they sort out repo. Let me know if it helped. ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-TJ98hAIN5io.html

You should be able to run without any of those settings. I see some use only --no-half-vae Check out this post for some discussion on RX6xxx settings github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/8626 Let me know.

Hi, Can you check if this applies to your problem? ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-TJ98hAIN5io.html

Same here, any solution?

Hi, are you using the correct version of Python. Also make sure you run the commands in admin mode command prompt window. Lastly some firewalls may cause timeouts on some of the repos. Let me know if you come right...

Hi, are you using the correct version of Python. Also make sure you run the commands in admin mode command prompt window. Lastly some firewalls may cause timeouts on some of the repos. Let me know if you come right...

Hi, Yep it has to match the prescribed version on the repo.

@@stuffaboutstuff4045 Thanks for this info. SD didn't work on my end though. Hope I'll get a proper card soon.

Hi, Download & install the required software, when installing remember to use the exact same version of Python (3.10.6), also execute all commands from "Admin" Command prompt, the rest should work if you follow the steps in the video. Let me know...

Hi, Maybe check out this link and let me know if it helps. www.reddit.com/r/StableDiffusion/comments/y2yoex/automatic1111_train_tab_guide_needed/?rdt=55796

Thanks for reaching out..

Hi, Looks like something is wrong with the model you downloaded. Make sure its not corrupt model. Also make sure you run all the commands used in the video in Admin Command Prompt. You also have to use the exact Python version specified in the video.. If you followed the steps and model is working then maybe check out this thread if still stuck. Everything should work if you follow the steps in the video but as a last resort looks at the link below.. github.com/AUTOMATIC1111/stable-diffusion-webui/issues/13087

Hi I will investigate this a bit. You should be able to run it reasonably fast with your spec GPU. Should be able to run it with no special settings in the config..

@@stuffaboutstuff4045 Thanks, I really hope you can help me, I manage to make it work but it´s really slow, it seems Stable Difussion is not made for AMD cards because of not CUDA support, or that´s what I read.

@@foreropa If you're on Windows I've heard another solution is to install Linux and then ROCm which supposedly should allow SOME AMD GPUs to work with CUDA. ROCm may work with some older AMD GPUs but some claim it's fine if you're working with an RDNA GPU. Running into roadblocks because of AMD GPU has been annoying. Also @stuffaboutstuff4045 thank you for the video. I'm excited to try this as I'd been looking into older version of Comfyui (newer version supposedly just moved to cuda12.1) since I thought I'd be unable to use A1111 with my machine specs.

@@chrysanth267Thank you very much for your answer. I have managed to make it work in windows, I don´t know if it´s slow or not, but well, at least it works.

remove the medvram option

Hi, what commandline_args are you using? You should have --medvram set. That is a 8GB card. Also check the link github.com/AUTOMATIC1111/stable-diffusion-webui/issues/5468

Hi, Glad to hear its working! Thanks for the feedback. Did you update the config file according to your GPU video memory. When I produce the image on an older RX580 8GB it takes 55 seconds using the Dreamshaper model. Not super fast but for what I use it for more than sufficient. I hope you get it sped up...

Yes updated the config file, everything works with v1-5. Thanks! Unfortunately it works not with SDXL 1.0 on my machine.

@@stuffaboutstuff4045 Thanks for that, i have an AMD 8GB dedicated video Ram and each picture is taking 5min to come out with 17s/it !! looks like my machine is ignoring the GPU and using only CPU... i have the --medvram config already but there must be another file to tell the stable diffusion to use the GPU rather than CPU.. Thanks

Hi, what command line arguments are you using. One of the users on the other video mentioned these worked well for him on a 6700XT, set COMMANDLINE_ARGS=--use-directml --medvram --precision full --no-half --no-half-vae --opt-split-attention-v1 --opt-sub-quad-attention --disable-nan-check. Did you check out the below link and try the switches under Automatic Installation heading. Let me know if this helped.. github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Install-and-Run-on-AMD-GPUs

@@stuffaboutstuff4045 ​ @stuffaboutstuff4045 Whenever I use those command lines its says io need "to ---skip-torch-cuda-test" so I plug it in and then is says I need to delete --use-directml and after I delete --use-directml its back to only using my CPU. and yes I have python 3.10.6 and git properly installed and my current command line is "--skip-torch-cuda-test --no-half --precision full --use-directml --medvram --precision full --no-half --no-half-vae --opt-split-attention-v1 --opt-sub-quad-attention --disable-nan-check" and it still gives me the prompt to delete --use-directml and when I do delete it doesn't use my gpu but when I have --use-directml stable diffusion wont boot and it will give me a error code

Hi, Did you check this video below. Is this not the issue you are sitting with? ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-TJ98hAIN5io.html This video should resolve your issue.

Hi, Did you check this video below. Is this not the issue you are sitting with? ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-TJ98hAIN5io.html This video should resolve your issue.

Hi, I had dodgy results using the .bat file or shortcut directly. Have been running out of CMD and everything works perfectly. Will look into this and revert. Thanks for reaching out!

Same.

Hey there! The same thing happened to me, this is what fixed the problem: -Clone SD again on a different folder. -Open and add on 'requirements_versions.txt' the following: torch-directml. -On webui-user add '--use-directml' on set COMMANDLINE_ARGS= -Run webui-user. It should continue with the install without problems.@@secundus6457

@@elementpricee Thanks for your help! But now I have: "AttributeError: module 'torch' has no attribute 'dml'"

Hi, Are you using the exact same version of Python when installing (3.10.6), also execute all commands from "Admin" Command prompt, the rest should work.

@@stuffaboutstuff4045 yeah

Hi, Did you check out this video. Is this the issue you have... ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-TJ98hAIN5io.html

Hi, If you followed this step by step it should work fine. Can you share the error.

@@stuffaboutstuff4045 Ok, I'm going to do the whole procedure again (I've reset my PC), if the error appears I'll share it here.

@@stuffaboutstuff4045 That's the mistake, right at the end: Installing requirements Launching Web UI with arguments: no module 'xformers'. Processing without... no module 'xformers'. Processing without... No module 'xformers'. Proceeding without it. Traceback (most recent call last): File "C:\stable-diffusion\stable-diffusion-webui-directml\launch.py", line 48, in main() File "C:\stable-diffusion\stable-diffusion-webui-directml\launch.py", line 44, in main start() File "C:\stable-diffusion\stable-diffusion-webui-directml\modules\launch_utils.py", line 476, in start import webui File "C:\stable-diffusion\stable-diffusion-webui-directml\webui.py", line 13, in initialize.imports() File "C:\stable-diffusion\stable-diffusion-webui-directml\modules\initialize.py", line 34, in imports shared_init.initialize() File "C:\stable-diffusion\stable-diffusion-webui-directml\modules\shared_init.py", line 26, in initialize dml.do_hijack() File "C:\stable-diffusion\stable-diffusion-webui-directml\modules\dml\__init__.py", line 70, in do_hijack if not torch.dml.has_float64_support(device): File "C:\stable-diffusion\stable-diffusion-webui-directml\modules\dml\backend.py", line 42, in has_float64_support return torch_directml.has_float64_support(get_device(device).index) File "C:\stable-diffusion\stable-diffusion-webui-directml\modules\dml\utils.py", line 7, in get_device device = torch.dml.current_device() File "C:\stable-diffusion\stable-diffusion-webui-directml\modules\dml\backend.py", line 48, in current_device return DirectML.context_device or DirectML.default_device() File "C:\stable-diffusion\stable-diffusion-webui-directml\modules\dml\backend.py", line 51, in default_device return torch_directml.device(torch_directml.default_device()) File "C:\stable-diffusion\stable-diffusion-webui-directml\venv\lib\site-packages\torch_directml\device.py", line 38, in device raise Exception(f"Invalid device_id argument supplied {device_id}. device_id must be in range [0, {num_devices}).") Exception: Invalid device_id argument supplied 0. device_id must be in range [0, 0).

Figured it out! Needed to install a compatible version of httpx

@@dlawrence How did you manage to make it work, the same thing happens to me.

@@Ayaplian8 Try typing these two lines: cd stable-diffusion-webui && source venv/bin/activate pip3 install httpx==0.24.1

Glad you hear you got this resolved.

Hi, The tips on getting everything working, Correct version of Python, run the commands in admin mode command prompt, make sure you have correct config setting for VRAM matching your GPU. If you still have problems let me know what AMD card you are using...

😉 lol neither. But the aussie guess is closer.

Hi, Mine used the GPU. It goes to about 80% and sits till complete. Takes 55 seconds per image on average using the Dreamshaper model.. Thank for reaching out..

Hi, just check if this applies to you did you check this SD video.. ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-TJ98hAIN5io.html

Hi, to get SD working currently you have to follow the original video but you also need to follow the steps in the updated video to remedy some issues. Can you check out the new video and let me know if that gets you up and running. ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-TJ98hAIN5io.html

Hi, Did you check that you are using the exact version of Python. Make sure you run the commands in admin mode command prompt. Let me know if problem persist?

Hi, Thanks for the input and for reaching out! Much appreciated. For those still stuck with --skip-torch-cuda-test. Please check out this video ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-TJ98hAIN5io.html

Hi, Make sure you have the correct version of Python as per video. Make sure you run all the commands in admin mode command prompt. Let me know if this resolves the issue..

Hi, The tips on getting everything working, Correct version of Python, run the commands in admin mode command prompt, make sure you have correct config setting for VRAM matching your GPU. If it still does not work let me know what video card you have.

@@stuffaboutstuff4045 I have a Radeon RX-480 8GB