Custom path for models

[system1system2]

(it's a question, but if the answer is "no", then it should be filed under "Ideas")

Is it possible to define a specific path to the models rather than copy them inside stable-diffusion-webui/models/Stable-diffusion?

Right now, I drop in that folder simlinks to my actual folder where I organize all my models, but it's still tedious. It would be easier to have a place where to specify your folder of choice and be done with it.

[CCpt5]

I've been hoping for this feature/option for a while also. I'd love to be able to move my models to a different drive with more space. When originally installing SD I didn't anticipate having so many models at 2GB each, but Dreambooth has created this situation.

Maybe this isn't possible?

Thanks for starting a thread on it.

[system1system2]

If you use Windows, you should be able to create a symlink for files in an external or alternative folder: https://skimfeed.com/blog/symbolic-links-in-windows-for-pointing-a-folder-to-another-folder-on-an-external-hard-drive-or-ssd/

I'm on macOS so I can't be of further assistance.

[CCpt5]

I'll look into this (hopefully as just a temporary solution). Thanks!

[Ypdes]

UPDATED. Symbolic links work well with checkpoints directory but don't work for loras, embeddings, hypernetworks etc directories in tabs below the prompt field. When I click the link, it opens the target directory, so the link is ok. But Automatic doesn't see models there, it shows "Error" instead (for ex. in LoRa tab). Any ideas?

So I'm still looking for the better approach, like in Fooocus - with separated file with all paths.

[onusai]

you can specify a models folder by adding --ckpt-dir "D:\path\to\models" to COMMANDLINE_ARGS= in webiu-user.bat

you can also do this for hypernetworks and embeddings
--hypernetwork-dir "F:\path\to\hypernetworks"
--embeddings-dir "F:\path\to\embeddings"

[rpillala]

It's also possible that it prefers local and if a model is not in the local directory it checks the one from the command argument. I already had my checkpoints on the NAS so it wasn't difficult for me to test moving them all and pointing to the NAS. What I haven't tried is moving just one back and see if it can find the others on the NAS with the command line argument active.

[system1system2]

I tried @rpillala and it works: some models are retrieved from the specified directory while others are retrieved from the local directory.

[mykeehu]

I use it on Colab: the models and settings are on Google Drive, so I mount it first, then clone and start. That's why the option to specify a unique VAE directory is missing, because I call everything else from Drive.

[boazcstrike]

This didn't work for me when I placed 'A:\path' and it reads Apathtomodels on the console. It worked when I replaced it with A:/path/to with the opposite slash. 😁 Hope this helps!

Here's my sample added .bat codes:

...
set COMFY_HOME=A:/ComfyUI
set VENV_DIR=%COMFY_HOME%/env
set COMMANDLINE_ARGS=%COMMANDLINE_ARGS% ^
 --ckpt-dir %COMFY_HOME%/models/checkpoints ^
 --hypernetwork-dir %COMFY_HOME%/models/hypernetworks ^
 --embeddings-dir %COMFY_HOME%/models/embeddings ^
 --lora-dir %COMFY_HOME%/models/loras
...

[cooperdk]

This didn't work for me when I placed 'A:\path' and it reads Apathtomodels on the console. It worked when I replaced it with A:/path/to with the opposite slash. 😁 Hope this helps!

Here's my sample added .bat codes:

...
set COMFY_HOME=A:/ComfyUI
set VENV_DIR=%COMFY_HOME%/env
set COMMANDLINE_ARGS=%COMMANDLINE_ARGS% ^
 --ckpt-dir %COMFY_HOME%/models/checkpoints ^
 --hypernetwork-dir %COMFY_HOME%/models/hypernetworks ^
 --embeddings-dir %COMFY_HOME%/models/embeddings ^
 --lora-dir %COMFY_HOME%/models/loras
...

On Windows, you must do double backslash because single backslash is an escape character meant for stuehuset character input.
Alternatively, forward slash as you mentioned.

[Eleksar387]

I had to use single quotes for the path --ckpt-dir 'E:\Stable Diffusion\Stable-Diffusion-Web-UI\Stable-diffusion\' to make it work (Windows)

[huizne]

Just want to add for any macos users. I had to use single quotes as well. Now I can store all my models in Google Drive. Thank you!

[OrganicBeej]

Hi there :)

I need to move Models directory to a separate 2TB drive to create some space on the iMac so followed these instructions for command line args. The Terminal window seems to show that A1111 has recognised the path but it started then to d-load a new models directory for SD (where the original one was in models/Stable-Diffusion) as I had backed this up to the 2TB drive and moved it to my Desktop, just to test to see if connected to the path?

Here is the Terminal start-up which I thought looked like it was going to work?

Launching Web UI with arguments: --skip-torch-cuda-test --disable-nan-check --upcast-sampling --no-half-vae --use-cpu interrogate --opt-sub-quad-attention --ckpt-dir \Volumes\BEEJ2TB\SD_Models\Stable-diffusion\ 

and here is the actual code from inside the web-user.sh file:

export COMMANDLINE_ARGS="--skip-torch-cuda-test --disable-nan-check --upcast-sampling --no-half-vae --use-cpu interrogate --opt-sub-quad-attention --ckpt-dir '\Volumes\BEEJ2TB\SD_Models\Stable-diffusion\' "

Would really appreciate help from a Mac OS user to let me know if I have done this wrong. Just to be clear, when starting A1111 it only had the basic 1.5 which was the one it downloaded in the stable-diffusion directory it created (because nothing was there at this time).
I did do a re-launch to see if it perhaps needed 'both' directories - one in the local install where it usually should be with the only 1.5 model - and one on my 2TB drive with ALL my other moved models.

Thanks for any help :)

[OrganicBeej]

😂🤪🤷🏼‍♂️I'm such an eedyot!! I just realised, as I watched it load again in the Terminal - ALL the directory slashes are backwards, which I guess must be correct for a PC but not right for an iMac! You'd think after using a Mac for 30-ish years I'd know better lol!

Oh well - old age a-kickin in I guess hehe.

Thanks for the post though, this has definitely worked now and Im currently rendering. Slower, so I guess maybe keep the most regularly used models in main iMac directory but just keep all the others in the 2TB. Nice :)

[zodchiy-ua]

I had to use single quotes for the path --ckpt-dir 'E:\Stable Diffusion\Stable-Diffusion-Web-UI\Stable-diffusion\' to make it work (Windows)

Finally got it working! Thanks man, you made my day! 🙏

[cooperdk]

Listen guys. There is exactly ONE way and most people failed to give the right answer.

Make a symbolic directory link in the webui/models/Stable-diffusion folder.

There is no other way to do this. Those who keep ranting about the ckpt-folder fail to understand that defining another folder will cancel out the original one. And you can only specify one!

Do it like this:

mklink /d <stable-drive-and-webui>\models\Stable-diffusion\e-drive-models e:\models

This will create a symbolic directory link named webui/mosels/Stable-diffusion/e-drive-models and it will point to the directory e:\models

[cooperdk]

mklink /d (brackets)stable-drive-and-webui(brackets)\models\Stable-diffusion\f-drive-models F:\AI IMAGES\MODELS
The system cannot find the file specified. <- here where

mklink /d d:\AI\stable-diffusion-webui\models\Stable-diffusion\F-drive-models F:\AI IMAGES\MODELS
The syntax of the command is incorrect.

what's wrong? Nothing works

You have a space in your directory name, so you have to refer to it in double quotes:

"F:\AI IMAGES\MODELS"

[muvo4k]

i dont know if docker is the right choice but i would say dont use an isolate environment especially when wrong dependencies can kill your running version and that is more likely that you need the 3secs. Especially consider when you have to troubleshoot that also time and i guess you can substract that but nobody does

[mikejk8s]

Containerizing these workloads has absolutely zero difference when it comes to performance throughput. Every single datacenter renting out GPUs has you running a container that mounts them via vfio directly to the operating systems cgroup calls. This is for performance & security. We don't want to see your data, other customers shouldn't see your data, and we want it off the server quickly. This used to be maybe be an argument with vms, a bad one usually, but the entire world has been running on containers since at before 2015, which is when I first got into distributed system workloads that are distributed via... containers.. VMs were a lot different and had shims between "harddisk" images and all that 9p network filesystem drama, in other words it just had layers of minute latency introduced through the path. Oci/containers are nothing like that.

I don't mean this as a brag but feel free to look at my github I've worked in container distsys for awhile and these workloads are completely stable and 100% expected to be containerized. Why would there be a CUDA Container Toolkit? Nvidia is/was heavily involved with many Kubernetes and container projects.

If stability matrix didn't work so well I was going to write up a docker-compose.yaml that would just launch all this stuff and let you version it by containers, a torchvision container, torch container, etc. Not having to pip uninstal and wait all that. After working on containers as much as I have, coming back to this "pip install reqfiles" style of deploying things feels like the 1970s.

Anyway, just figured I'd chime in and refute that myth.

[cooperdk]

Containerizing these workloads has absolutely zero difference when it comes to performance throughput. Every single datacenter renting out GPUs has you running a container that mounts them via vfio directly to the operating systems ...
Anyway, just figured I'd chime in and refute that myth.

I would argue that there is a world of difference between datacenter containers and docker containers run from a desktop pc.
I know many CPIs support hypervisor and other VM platforms but it's two different scales and price ranges - and therefore end results. I have trouble believing that a docker containers is as fast as running the same outside of it. Especially when the GPU is involved. Never tested it though.

[mikejk8s]

@cooperdk Let me try this again.

"Hi, I literally work on this specific, exact, precise thing that we're discussing that you have mentioned that you're unfamiliar with and (as posted here) "never tested it though." and this is the answer to the hypothesis of which you haven't tested. Because I have tested it. At many companies. In many environments.

you: "I would argue X, Y, Z..... never tested it though."

I'm not asking for input, I'm telling you the factual answer. Containerizing these workloads has been done for decades and is not a performance issue or concern whatsoever in any capacity with qemu/vfio/cgroup hooks. There is not a single massive datacenter out there not running these gpu workloads in containers.

We ALL containerize them. I am intimiately familiar with nvidias workloads.

You clearly do not understand what a "docker container" is. Which is fine. If you'd like to dive into that we can, but in the shortest sense, no, there is no performance penalty for running processes inside of cgroup/oci containers. The os.syscall hooks are are managed by cgroups and not any shims that would add additrional latency.

The hardware is directly passed thru via vfio and the host OS know nothing about it aside from some PCI UUIDs it mounts between container and bios.

I understand this is a stablediffusion repo so there are a lot of people no knowledgeable about distributed systems in here, that's also fine. But keep in mind, you're now on github with actual programmers and contributors. This isn't reddit. Some of us are involved that low.

Please, as someone who has worked in containers and disibributed systems for a very long time, as my actual engineering job, stop perpetuating the "VMs slow things down" myth. The ENTIRE world runs on qemu vms nowadays. Every single EC2, GCP, DigitalOean,Heroku,blahblah are all qemu vms. I've been in operations/infra for almost 20 years and even back then we used Solaris Jails to the same affect.

And no, there is literally no difference between an OCI container on your desktop or my $150k server. You have no idea what a container actually is. It's a process. That's it. It could be explorer.exe running in its own CGROUP which means its in its own CONTAINER. It is just a PROCESS, an x86 process runs like an x86 process on any x86 processor, the only difference is we have meta data around it that tells the OS to isolate it into its own cgroup.

[Lycoris-SF]

Guys I cannot do it on extensions folder:

[Lycoris-SF]

DominikDoom/a1111-sd-webui-tagcomplete#132 (comment)

[rbs250]

Is there a way to add LoRA and controlNet folders as well?

[cooperdk]

Same way

[noisefloordev]

Listen guys. There is exactly ONE way and most people failed to give the right answer.

Super toxic

[robert-the-slayer]

Ok , here is example of how cool is Linux like operating systems. Just put symbolic link into any kind of folder that you want to expand. For example :
cd (path_to_ automatic1111)/models/Stable-diffusion
ln -s /media/USB (or whatever you need) checkpoint_models_on_usb

That all folks!
Now just run your a1111 , or refresh models list in dashboard and Enjoy ;)

P.S. you could create for any , LORA and etc , and as many as you need :)

[StanleyPain]

No one cares which operating system you use. Use whatever works for you.

[cooperdk]

Don't know what are you talking about. I am not making decision only because of one reason. I compared both systems in different ways , observed different features and sides , Linux wins in literally all possible points , for me. I never come back to Windows , Windows is scrap. And i am so happy that world slowly goes into Linux sector. Windows is greedy on money and device resources , buggy and laggy piece of crap , ugly. Thank you God and heaven , that there are such a great systems family like Linux. Edem exist on Earth and name of it is Linux. I am regret only one thing in my life , that i was in Windows too much , and spend few dollars on it :)

It does not win with stable diffusion, no matter what the f##k you say, Linux is noticeably slower. I don't care about anything else in this thread.
Ps you can run pure Linux inside Windows in numerous ways. That way you can use Linux AND use an OS for games and software that's actually supported.

[Ainaemaet]

There are definitely some advantages to sticking with windows if that is what you use most of the time, but If windows is _much faster than WSL2 or another distro for your gens, you don't have it set up right .

I use WSL because a large number of ML functions don't have a windows equivalent, or what is available isn't very good - and if you are running a lot of cutting edge stuff it is a huge godsend to get used to it.

I get faster gens in my WSL setup, but it took me a while to get things figured out - and Linux certainly isn't for everybody.
Certainly not as 'user friendly' as windows, but far more power user friendly in many ways.

Edited to say a couple years instead of a few because I first came to the thread in 2022, but I'm getting ahead of myself and thought more time had passed!!! xd

Many of the people I have spoken to in 4090 grotto and a few other places have also said they are doing better running in various Linux distros than windows - I would imagine a good deal of that owing to the gain in overhead (Linux doesn't load even half the crap and takes up much less memory than windows).

It seems odd to me that you would get upset about it; often when I see that my first assumption is that the person who is commenting is not really as accustomed to Linux as they make themselves sound, and so there is a (quite understandable) bit of a knee-jerk reaction in feeling challenged that they are out of their element there.

I don't think many Linux pros would call any version/distro running inside windows 'Pure', regardless of whether it's in a VM or otherwise (is there something aside from VM/WSL/Docker? Please let me know because I haven't heard of it and would love to check it out!!!).

Anyhow, I'm not here to argue or get anybody riled up - I just stumbled on the post because I had been a part of the discussion a couple years ago and thought perhaps I could help some of those (and possibly yourself) who might see your comment and assume Linux won't work as well for stable-diffusion to see that is definitely not (necessarily) the case!!! :)

[Ainaemaet]

I can't help but think some of the downvotes (I had the same thing happen to me a couple years ago when the thread was fresh) are from people who just aren't used to Linux (it certainly isn't exactly 'user friendly' for most people, even if your using something with a good GUI lol).
Also, most people here are going to be gaming and doing a lot of other things other than machine learning, and while Linux is far ahead of where it used to be for games, it's still nowhere close to what windows provides.

I run all of my ML programs in WSl2 anaconda, and I can 100$ confirm that even there I get faster gens than my windows install - BUT it took me a lot of time and effort to get things this way so I can't fault others for not agreeing with you lol. XD

I prefer windows by far over linux for almost everything that isn't unsupported and I don't have any reason for.

[cooperdk]

There are definitely some advantages to sticking with windows if that is what you use most of the time, but If windows is _much faster than WSL2 or another distro for your gens, you don't have it set up right .

other places have also said they are doing better running in various Linux distros than windows - I would imagine a good deal of that owing to the gain in overhead (Linux doesn't load even half the crap and takes up much less memory than windows).

Believe me, I had it setup right. I have worked on many levels in the it sector, including administrator in a company that did laboratory tests on soil etc, and where the hardware was very sensitive.

Fact is, the gpu drivers are nowhere as mature on Linux as on Windows.
I did a test on the same Cuda release with the same release of webui.
I then did a test using the same prompt with identical settings on Windows and Linux. I setup the number of steps so it would run at least a minute on my RTX 3060.

I even posted the results here with a screenshot, I believe, because the difference was big.
Like at least five seconds, Linux being the loser.

And here is the thing, commenting on the second snippet of your reply:
I specifically installed Linux just for this test. No other software was installed. The Windows release used in the test had a number of typical apps installed, although it was what I would call a new install, this should actually have been in Linux' favor. It wasn't.

BTW, I have 64 GB RAM and my 3060 is the 12 GB version.

I guarantee you this test was fair and in fact in Linux' favour.
My guess as I said is that a lot of Python packages are actually made for Windows, and while the AI ones are probably not, it does not help because the drivers are still somewhat... more like, not at all mature and likely in beta or even alpha stage when it comes to Linux. Does not help that Linux is generally fast if it's drivers aren't. And yes, I did install everything correctly.

This was a year ago. I doubt the result will be different now. My test install of Linux? Deleted the partition. I had good it would be faster as I would then have used it for AI. Turns out using Linux for this is not a good idea at this point.

PS: I have the screenshots of the results at home, will try to remember to post them when I am there.

BTW: if you run your AI in WSL, it by definition cannot be faster than on your Windows. That's how visualization works.

I agree with you on one point: It is kind of easier to set up ML/AI Git projects on Linux because on Windows there are a lot of prerequisites. That would be my only reason to use Linux at this point.
(BTW I tested it in WSL too, but the above test was actually run as main OS from a NVMe drive). It was, by the way, a Debian based release for which "optimized" GPU drivers were supposed to exist... (Linux Mint)

[enternalsaga]

Is it possible to set up a secondary/third models directory? I'm running SD on colab so I'd like it to read the colab folder and my local drive as well.

[cooperdk]

Not colab unless you have your models on Google drive and mirror it on a local drive... but you can map any local drive using the regular command for your system (Windows: mklink). Your map links must be inside the webui models directory.

[maniac-0s]

I have the same issue. I am also running EasyDiffusion (and also want to try Comfy UI sometimes). It's quite inconvenient that I can't set a models folder. I tried to create a symlink but A1111 will just create a new models folder and claim it can't find anything in there.

I don't want to copy 100 of GB of models and loras etc to every UI that I use. I will go with the comandline option but I don't find it in the bat. Does it matter where the variable is placed in the file?

[StanleyPain]

Use mklink.

https://www.howtogeek.com/16226/complete-guide-to-symbolic-links-symlinks-on-windows-or-linux/

I have model and output directories shared that way across webui, comfy, foocus....

[cooperdk]

There is already a long issue thread covering this where I have provided software-specific examples.

[maniac-0s]

Symlink is of course a workaround but not a fix to the feature request. Same as using comandline option. But as I understand it was asked if we could have this as a settings. A1111 can specify output folders, why not input folders (read: models)?

[noisefloordev]

Use mklink.

https://www.howtogeek.com/16226/complete-guide-to-symbolic-links-symlinks-on-windows-or-linux/

I have model and output directories shared that way across webui, comfy, foocus....

I've been doing this for a long time. It works, but it's a pain since it confuses git. I have to undo it all to update, and then redo it after. (In principle git stash should be able to do this, but it doens't seem to work in this case in Windows.) But it's worth it so my data is in a clean, centralized place and not interleaved with the application.

I wish you could just set a user data directory where all of your mutable data lives (not just checkpoints, but all models, temporary files, settings, hypernet/embedding training output, etc.), and have the A1111 installation itself be read-only like most applications. That would keep the application tree clean, and also allow having multiple installations pointing at the same data, so you could test a new branch or release with your existing data without disturbing the version you know works.

[maniac-0s]

this

[StanleyPain]

I have never had git become "confused" by this and that's with 3 different webuis running off the same data.

[cooperdk]

I have never had git become "confused" by this and that's with 3 different webuis running off the same data.

It doesn't get confused. Git does not even consider these directories. The user above has done something wrong.