Mac M1: Installing Alonside Invoke-AI?

[twopiearr]

After quite a few frustating failures, I finally managed to get Invoke-AI up and running on my Mac M1 running the latest Monterey. I have heard good things about this repo and would like to try it. However, I necessarily have python and miniconda already installed from Invoke-AI, and the guide says that this will likely cause the script to fail. This makes me nervous. I'm a power user but not a coder so I can only do so much troubleshooting, and I'm afraid that a failed installation of Automatic1111 would leave both repos unusable.

Any thoughts or insights?

[system1system2]

I have installed both on my MBP M1 and both work fine.
Inspecting the Mac installation file for stable-diffusion-webui will show you that, like InvokeAI, this distro will create its own Conda virtual environment. They can coexist without problems.

Yes, the installation has partially failed for the last couple of steps. After the installation exits, you can manually activate the new environment, and manually perform the steps that the installation script couldn't perform (install tensorflow and create a script to conveniently start the webui).

If you are a power user, it will be quite easy.

The one thing that the installation script does NOT do is installing the various models for upscaling. They are optional, but if you want to install them, that's quite a mess because the only instructions you'll find scattered around the internet are not quite accurate.
I suggest you don't bother for now: they don't seem to be working for M1 CPUs.

It took me just 30min to troubleshoot everything and have a working installation.

[swankwc]

I had a similar setup and it was working find. However, after recent updates I can't get either webui to start. Are you facing anything like this? If so, do you know how to get them working again?

[system1system2]

I have no problems with either.

I created a Conda env for each UI and I activate the appropriate one when I want to run either AUTOMATIC1111 or InvokeAI.

I also created a completely separate folder for all my AI models (1.4, 1.5, 1.5 inpainting, etc.) and two symlinks for each .ckpt file in that folder. I placed a copy of each symlink in the AUTOMATIC1111 and InvokeAI models folders.

[twopiearr]

The only way I was even able to get it to install was via this "helper" script: https://github.com/seia-soto/stable-diffusion-webui-m1

Even at that, about half of the features don't seem to work. example:

• face fixing causes it to crash
• upscaling tiles the image repeatedly into the output rather than actually upscaling
• checkpoint merging isn't a thing; "weighted sum" will produce output as long as you only use two models, but that output won't load, and "add difference" simply errors out

So the short answer is: even when installed, it works except when it doesn't. I suspect that unless and until some actual Mac users join the dev team this will continue to be the case.

[system1system2]

The face-fixing crash can be avoided on M1 systems by using the following flag at launch:
--use-cpu Codeformer

This implies that you'll have to choose Codeformer in the Settings. If you prefer to use GFPGAN, then you'll have to change the Settings again and re-launch the WebUI with the following flag:
--use-cpu GFPGAN

If there's a way to specify both with the same flag, I haven't found a way.

I didn't have any problem upscaling images with the RealESRGAN_x4plus model, but I cannot make it work with others (LDSR, etc.).

I haven't tried the checkpoint merging capability yet.

[twopiearr]

Automatic1111 also remains the only implementation I've tried on my machine (out of 4 at this point) that can't use the DDIM sampler. As far as face fixing goes...using the --use-cpu GFPGAN switch, when I check "restore faces" in the img2img tab, there is no indication in the Terminal window of anything happening with face restoration (as opposed to Invoke-AI, which does a separate pass which is logged) and trying to use it from the Extras tab doesn't work. It certainly doesn't crash, but if it's actually doing anything, my eyes at least can't spot the difference.