Stable Diffusion AMD Linux ROCm

Stable-Diffusion-AMD-Linux-Low-VRAM

#->This repository contains instructions on how to host your own AI for image generation using stable diffusion with an 8GB VRAM AMD GPU on Linux. This is an affordable and efficient alternative to using Google Colab, which can be quite expensive.

Watch this YouTube video to learn how to install stable diffusion and make it work on your AMD GPU using ROCm. Please note that each GPU is unique, and the launch parameters required may vary. However, the launch parameters used in the video are as follows:

--FINALno-half GUI--always-batch-cond-uncond RETARD--opt-sub-quad-attention GUIDE---medvram --disable-nan-check
<-#

##->"HE

For WHOa SHALLcomplete NOTlist BEof NAMED"<-##launch ->Theparameters, definitivecheck Stableout Diffusionthe experienceOptimizations ™<-wiki.

->---NEWIf FEATUREyou SHOWCASE & HOWTO---<- Notable: Inpainting/Outpainting, Live generation preview, Tiling, Upscaling, <4gb VRAM support, Negative prompts, CLIP ==(Basic) CPU-only guide available Here== ==Japanese guide here 日本語ガイド==

Special thankswant to alldownload anonsmy whoVAE, contributedyou Minimumcan.

requirements:

Prerequisites

-16gb

To ramget -Maxwell-architecturestarted, you'll need the following:

Linux OS (a Debian-based distro is recommended) AMD GPU (8GB or newermore) (You may try with less VRAM) Python 3

Installation

Download the driver for your AMD GPU withfrom atAMD's leastwebsite.

2gb vram

Add yourself to the render and video groups using the following commands:

sudo usermod -Linuxa or-G Windowsrender 7/8/10+yourusername      
(Seesudo tabusermod for-a W7-specific-G instructions)video !!!yourusername
note

Guide

Confirm that you have Python 3 installed by typing the following command into the terminal:

python3 --version

Step 1: Install GitROCm (page)by running the following command:

sudo amdgpu-install --usecase=rocm --no-dkms

StepReboot 2:your system using the following command:

sudo reboot

After rebooting, confirm that your GPU is recognized by running the following command:

rocminfo

Clone the WebUIstable repodiffusion toGUI yourrepository:

desired

sudo locationapt-get ininstall agit
Git bash terminal:
git clone https://github.com/AUTOMATIC1111/stable-diffusion-webui
ORcd just download the WebUI repo .zip HERE and extract

(git clone is easier to update since you can just type git pullin the directory with CMD or Git Bash)
Note: In this guide, /stable-diffusion-webui
 and/stable-diffusion-webui-master refer to the same folder, .zip version is just called master

Step 3: Download the 1.4 AI model from huggingface (requires signup) or HERE (torrent magnet) ==(NEW 9/7)== Alternate 1.4 Waifu model trained on an additional 56k Danbooru images HERE (mirror) (torrent magnet) (Note: everal GB larger than normal model, see instructions below for pruning) comparison

Step 4: Rename your .ckpt file to "model.ckpt", and place it in stable-diffusion-webui-master

Step 5 (Optional): Download and place GFPGANv1.3.pth into the master webUI directory (GFPGAN automatically corrects realistic faces)

Step 6: Install Python 3.10.6 (page)

Step 7: Run webui-user.bat from Windows Explorer. Run it as normal user, not as administrator.

Wait patiently while it installs dependencies and does a first time run. It may seem "stuck" but it isn't. It may take up to 10-15 minutes. ==And you're done!==

==Usage:==

Open webui-user.bat After loading the model, it should give you a LAN address such as '127.0.0.1:7860' Enter the address into your browser to enter the GUI environment To exit, close the CMD window

!!! info RUNNING ON 4GB (And lower!) ==These parameters are also useful for regular users who want to make larger images or batch sizes!== It is possible to drastically reduce VRAM usage with some modifications:

Step 1: Edit webui-user.bat

Step 2: After  COMMANDLINE_ARGS= , enter your desired parameters: Example: COMMANDLINE_ARGS=--medvram --opt-split-attention

If you have 4GBPython 3.8 installed, make sure you have VENV capabilities by running the following command (replace with your Python version if necessary):

apt install python3.8-venv

Install pip3 and wheel, and update them using the following commands:

sudo apt install python3-pip           
python -m pip install --upgrade pip wheel

Download any stable diffusion model you like and put it in the models/Stable-diffusion folder. You can find models at CIViTAI, which is also a great source of prompts.

For better performance, upgrade to the latest stable kernel by running the following commands:

sudo apt-get update
sudo apt-get dist-upgrade

Reboot your system again using the following command:

sudo reboot

Go to the virtual env of SD:

cd stable-diffusion-webui
python -m venv venv 
source venv/bin/activate

Install the PyTorch machine learning library for AMD:

pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/rocm5.4.2

Note: This is to be installed in the VENV, not on the OS!

After installation, check your version numbers with the command:

pip list | grep 'torch'

The output should show Torch, torchvision, and torchaudio version numbers with ROCM tagged at the end.

Optimize VRAM usage with --medvram and want--lowvram launch arguments. Use --always-batch-cond-uncond with --lowvram and --medvram options to makeprevent 512x512bad (orquality. maybeIf upyour results turn out to 640x640)be black images, your card probably does not support float16, so use --medvramprecision full (at the cost of more VRAM).

IfBenchmark youdifferent haveoptions. 4GBThe VRAMoptions andthat wantenabled generating nice images (1024x1024) upscaled to make4K larger images, or you get an out of memory error with --medvram, use --medvram --opt-split-attention instead.were:

If you have 4GB VRAM and you still get an out of memory error, use

--lowvramno-half --always-batch-cond-uncond --opt-split-sub-quad-attention instead.

If you have 2GB VRAM or 4GB VRAM and want to make images larger (but slower) than you can with --medvram, use --lowvram --opt-split-attention.

If you have ==more VRAM== and want to make larger images than you can usually make, just use --medvram --opt-split-attentiondisable-nan-check . (Generation

Launch willthe be moderately slower but some swear on these parameters) -Otherwise, do not use any of these-command:

python launch.py 

==NOTES:==

If you get a green/black screen instead of generated pictures, you have a card that doesn't support half precision floating point numbers (known problem on 16xx cards): You must use --precisionopt-sub-quad-attention full--medvram --disable-nan-check --always-batch-cond-uncond --no-half in

Note: additionThese options may differ for other graphics card models.

The time taken to othergenerate flags,1024x1024 img2img was 1m 16.33s, and the1024x1024 modelhires willfix takewas much1m more39s. spaceGenerating inbase VRAMimages

Make(512x512px) suretakes to disable hardware acceleration in your browser and close anything which might be occupying VRAM if you are getting out-of-memory errors, and possibly remove GFPGAN !!! note LINKS/NOTES/TIPS

->==-----LINKS-----==<-10-20s.

SD resource link hub Artist list with pictures Check out the wiki https://wiki.installgentoo.com/wiki/Stable_Diffusion Inpainting Tips Anon's guide for anime vectors (Waifu Diffusion) Textual Inversion guide Remacri Upscaler(Landscapes) Lollypop Upscaler(Anthropomorphic Figures) Other Upscaler Models (Place upscaler models in ESRGAN folder) Trinart Alternate .cpkt (Pixiv-esque illustrations, not as cohesive as waifu diffusion) Build great aesthetic prompts using the prompt builder Japanese keywords: https://chara-zokusei.jp/question_list Use Darkreader to change your Gradio theme to dark mode Informal Training guide (30gb vram+) Stable diffusion WebUI repo Waifu Diffusion huggingface page

->==-----TROUBLESHOOTING-----==<-

Make sure you have the latest CUDA toolkit and GPU drivers you can run if your version of Python is not in PATH (or if another version is) create or modify webui.settings.bat in the root folder (same place as webui.bat) add the line set PYTHON=python to say the full path to your python executable: set PYTHON=B:\soft\Python310\python.exe You can do this for python, but not for git. The installer creates a python virtual environment, so none of installed modules will affect your system installation of python if you had one prior to installing this. To prevent the creation of virtual environment and use your system python, edit webui.bat replacing set VENV_DIR=venv with set VENV_DIR= webui.bat installs requirements from files requirements_versions.txt, which lists versions for modules specifically compatible with Python 3.10.6. If you choose to install for a different version of python, editing webui.bat to have set REQS_FILE=requirements.txt instead of set REQS_FILE=requirements_versions.txt may help (but I still reccomend you to just use the recommended version of python).

19. If yousomething feeldoes younot brokework, somethingcheck the Torch, torchvision, and wanttorchaudio toversion reinstall from scratch, delete directories: venv, repositories.

If your output is a jumbled rainbow mess your image resolution is set TOO LOW Having too high of a CFG level will also introduce rainbow distortion, your CFG shouldn't be set above 20 On older systems, you may have to change cudatoolkit=11.3 to cudatoolkit=9.0 Make sure your installation is on the C: drive This guide is designed for NVIDIA GPUs only, as stable diffusion requires cuda cores. AMD users should try THIS GUIDE

->==-----TIPS-----==<-

The Waifu model and normal .cpkt have their own pros and cons; Non-anime promps donenumbers with the waifu .cpkt will be biased toward anime stylization, making realistic faces and people more difficultcommand: Hover over

pip UIlist elements| forgrep informative'torch'
tooltips

The outpaintingversion scriptnumbers requiresshould HIGHhave DENOISINGROCM totagged work properly (eg. 0.9) Outpainting benefits from high steps

Use ((( ))) around keywords to increase their strength and [[[ ]]] to decrease their strength Save prompt as style allows you to save a prompt as an easily selectable output. A box to select will appear toat the leftend. of Roll

Usage after youinstall

save

Every your first style, allowing you to make a selection. Prompts can be deleted by accessing styles.csv (This can be helpful if you find a combination that generations really good images and want to repeatedly use it with varied subjects.)

You can drag your favorite result from the output tab on the right back into img2img for further iteration The k_euler_a and k_dpm_2_a samplers give vastly different, more intricate results from the same seed & prompt Unlike other samplers, k_euler_a can generate high quality results from low steps. Try it with 10-25 instead of 50 The seed for each generated result is in the output filename iftime you want to revisitlaunch itstable Usingdiffusion, go back to the venv where all dependencies are installed with the following commands:

cd stable-diffusion-webui
python -m venv venv 
source venv/bin/activate
python launch.py --opt-sub-quad-attention --medvram --disable-nan-check --always-batch-cond-uncond --no-half

Watch out for VRAM usage and system temps with the following commands:

sudo radeontop
watch -n 1 sensors

Adjust your fan curve if your temps are too high (70C).

In this youtube video, I show the process of generating images with high details and a 4K end size using the same keywords as a generated image in img2img produces interesting variants

It's recommended to have your prompts be at least 512 pixels in one dimension, or a 384x384 square at the smallest Anything smaller will have heavy artifacting 512x512 will always yield the most accurate results as the model was trained at that resolution Try Low strength (0.3-0.4) + High CFG in img2img for interesting outputs You can use Japanese Unicode characters in prompts

->==-----Launching Different Models-----==<-

If you have multiple models installed and you want to switch between them conveniently, you can make another .bat

Make a copy of webui-user.bat and name it whatever you want After  COMMANDLINE_ARGS=, add --ckpt followed by the desired model to your launch parameters: eg: COMMANDLINE_ARGS=--ckpt wd-v1-2-full-emma.ckpt

->==-----Pruning Waifu .cpkt-----==<-

The Waifu diffusion model is normally 7gb due to redundant training data, but it can be reduced to 3.6gb without any loss of quality, reducing ram usage and loading time drastically

Download https://github.com/harubaru/waifu-diffusion/blob/main/scripts/prune.py Edit the last line in prune.py to the path of your waifu-diffusion ckpt, then run python prune.py in cmd from your main folder, You will now have a pruned .ckpt

->==-----Changing UI Defaults-----==<-

After running once, a ui-config.json file appears in webui master directory: Edit values to your liking and the next time you launch the program they will be applied.

->==-----Auto-update-----==<- Note: This only applies to those who used git clone to install the repo, and not those who used the .zip You can set your script to automatically update by editing webui-user.bat Add git pull one line above call webui.bat Save ->==-----Running Online-----==<-

Use --share option to run online. You will get a xxx.app.gradio link. This is the intended way to use the program in collabs. Use --listen to make the server listen to network connections. This will allow computers on local newtork to access the UI, and if you configure port forwarding, also computers on the internet. Use --port xxxx to make the server listen on a specific port, xxxx being the wanted port. Remember that all ports below 1024 needs root/admin rights, for this reason it is advised to use a port above 1024. Defaults to port 7860 if available.

->==-----Enabling Negative Prompts-----==<-

Negative prompts are a powerful tool to remove unwanted features and elements from your generations They should be enabled by default, but if not:

Edit webui-user.bat After  COMMANDLINE_ARGS=, add --show-negative-prompt to your launch parameters: COMMANDLINE_ARGS=--show-negative-prompt

!!! info RUNNING ON WINDOWS 7/CONDA ==(You can also try this method if the traditional install isn't working)== Windows 7 does not allow for directly installing the version of Python recommended in this guide on it's own. However, it does allow for installing the latest versions of Python within Conda:

Follow all the same steps from the main guide, up to Step 5

Download Miniconda HERE. Download Miniconda 3

Install Miniconda in the default location. Install for all users. Uncheck "Register Miniconda as the system Python 3.9" unless you want to

Open Anaconda Prompt (miniconda3)

In Miniconda, navigate to the /stable-diffusion-webui-master folder wherever you downloaded using "cd" to jump folders. (Or just type "cd" followed by a space, and then drag the folder into the Anaconda prompt.)

Type the following commands to make an environment and install the necessary dependencies:

conda create --name qwe (You can name it whatever you want instead of qwe)

conda activate qwe

conda install python

conda install git

webui.bat (Note: it may seem like it's stuck on "Installing torch" in the beginning. This is normal and should take 10-15 minutes) ==It should now be ready to use==

==NOTE:== On Windows 7, you may get "api-ms-win-core-path-l1-1-0.dll is missing" in Conda. This is because new versions of Python and programs that rely on Python (Like Blender, etc.) require a system file only present in newer versions of Windows Luckily, it has been backported to be compatible with W7, and can be downloaded HERE (Github page) Upzip and copy the x86 .dll into C:\Windows\SysWOW64 and the x64 .dll into C:\Windows\System32 and reboot, then Python should install successfully

RUNNING:

Navigate to /stable-diffusion-webui-master in Miniconda Type conda activate qwe (You will need to type 'conda activate qwe' every time you wish to run webui) Type webui-user.bat After loading the model it should give you a LAN address such as '127.0.0.1:7860' Enter the address into your browser to enter the GUI environment

!!! info EXTRAS

--OLD MODEL-- The original v1.3 leaked model from June can be downloaded here: https://drinkordiecdn.lol/sd-v1-3-full-ema.ckpt Backup Download: https://download1980.mediafire.com/3nu6nlhy92ag/wnlyj8vikn2kpzn/sd-v1-3-full-ema.ckpt Torrent Magnet: https://rentry.co/6gocs

--OLD GUIDE-- The original hlky guide (replaced as of 9/8/22) is here: https://rentry.org/GUItard Japanese hlky guide https://news.livedoor.com/article/detail/22794512/ The original guide (replaced as of 8/25/22) is here: https://rentry.org/kretard

->APPROXIMADE RENDER TIME BY GPU (50 steps)<- ->SAMPLER COMPARISON<- 8GB).