How to load more models?

[willow lava]

So, I got invokeAI now started and loaded, but I only had one model downloaded during installation. Now I have copied many more models into the folder where the ckpts go, but even after starting it again in terminal, it still only loads the one model from the first run. How can I show the software that there are more models to load?

[willow lava]

uuuuuuhhhhhhh, so, I found more info here: https://github.com/invoke-ai/InvokeAI/discussions/711

But it is rather complicated, because I need some config (.yaml) file and a vae checkpoint file, but usually custom models don't have such stuff attached - my self-trained DB stuff so far only is one large ckpt file, so I don't know adding it manually like this would work

[brave sable]

You can start invoke and use the command line

and then use !import

which will handle the process for you

[willow lava]

Thanks, but this isn't specific enough. After starting invoke, the Terminal window is busy. Where do I type " !import" ? The console in the Web UI doesn't seem to accept commands, apparently only displays output?

[plucky ruin]

start it without --web (just python ./scripts/invoke.py), then you'll see invoke>

[willow lava]

Oh, so I need to choose 1 or 3 when starting the script so I get to a command line?

[plucky ruin]

Sorry, I didn't understood. What "1 or 3" do you mean?

[willow lava]

(base) Computer:invokeAI username$ ./invoke.sh --web
WARNING: overwriting environment variables set in the machine
overwriting variable {'PYTORCH_ENABLE_MPS_FALLBACK'}
Do you want to generate images using the

1. command-line
2. browser-based UI
3. open the developer console

the first thing I see when I run it

[plucky ruin]

ah, the new installer (I haven’t got one :-))

1

[willow lava]

okay, trying this

This seems to work, will import a few models now before trying to look for them in the WebUI

Thanks for now! 😄

[willow lava]

Sooo... I managed to import a few models and loaded the SD1.5 inpainting model. I fire up the web UI and go to inpainting. But when I try to import a (high resolution) image, the console outputs "Disconnected from server" for a moment, loads 0 images, then shows "Connected" again. Do I have to scale down the image first? Are 10 Megapixel too much?

Damn, loading a 512x512 image works...

Okay, it seems resolution is not the problem, but it has to be square format

[brave sable]

shouldn't have to be square 🤨

if you do something too big for your computer though, it'll fail.

a 10megapixel file is probably too big for inpainting lol

whats the resolution on it?

[willow lava]

3888 x 2592 Pixel. I assumed 😅 that the software would ignore any pixels outside the inpainting/masking box and I would be fine even if the rest of the image is large, as long as the inpainting box isnt too big

[brave sable]

it does, when generating, but the UI might not handle something that big... IDK

i havent tried - but could!

[willow lava]

Thanks! 🙂

[potent patio]

You are correct about how it works, and it doesn't need to be square - must be a different issue. If you have some time to troubleshoot I'd like to figure out what is going on.

A descriptive traceback error should be in the terminal window when something has trouble. Could you try to recreate the problem please and then copy-paste the terminal traceback here?

Ty

[willow lava]

Sure, I like to help! The terminal does not produce any new output when this problem occurs. All I can see is the output of the console in the web UI, which shows this:

2022-11-19T16:07:36+0100:
Disconnected from server
2022-11-19T16:07:37+0100:
Loaded 0 images
2022-11-19T16:07:37+0100:
Loaded 0 images

Those three lines are displayed every time I drag that image to the browser window

By the way, it seems that invoke uses pytorch and only MPS as a fallback? Isn't mps (used by Diffusion Bee) faster/better? I am looking especially for a RAM saving method, because my 16 GB MacBookPro (M1Pro) is doing a lot of swapping, and I don't want SD to kill my SSD

[potent patio]

Do you mind sharing the image here? It needs to be the exact image, not compressed or changed in any way, and I can do more in depth troubleshooting. If not comfortable with that, no worries I understand.

[potent patio]

PyTorch uses MPS with CPU as a fallback for certain operations which are not yet implemented on MPS. DiffusionBee uses an entirely different macOS-only ML backend. We haven’t thought about that yet because invokeai is cross platform.

[willow lava]

Let me look for a picture I can share which causes the same problem...

[willow lava]

Thanks for the clarification! 🙂

[past abyss]

I copy a few custom models that do not appear in the menu, I try to use them!import but it not work