n00mkrad / text2image-gui

Somewhat modular text2image GUI, initially just for Stable Diffusion
GNU General Public License v3.0
921 stars 93 forks source link

Two details: InvokeAI somehow less memory efficient and Model "double-conversion" #76

Open PhSMu opened 1 year ago

PhSMu commented 1 year ago

So, As mentioned in the title, the new 1.9.1 works again for the latin/international people. There are two little details which I need to make the dev aware of (and this is in no way a complaint about the tool):

  1. Invoke AI is somehow less memory efficient than the previous known good version (for me, I got the ., bug) 1.7.2, which again, allowed me 640x960 generations at he start of the tool, and once it 'warmed up', 960x960 or 640x1280, on a measly GTX 1660 Mobile Max-Q (yes, notebook user here), at the VRAM usage cost of 6136/6144. This version of Invoke AI in 1.9.1 breaks the moment I step over 512x576, showing the '0 bytes left' despite the 1660 reporting 6088/6144 in the worst case scenario. I'm not a programmer by any means, but I think that may be a bug.

  2. Model conversion works great, and the tip from another post to download .SAFETENSORS files and convert them to .CKPT to prevent the state_dict error is good, but there's a little problem with it, which is some files not being available in .SAFETENSORS format, and of course, if the program can't use the CKPT to begin with, it can't convert it to anything else either. And after using another program to convert said CKPTs to .SAFETENSORS and then back, they work without issues, even in 1.7.2. So I think a checkbox to ignore security errors in the conversion might help address that, and it would be a nice quality of life to have a CKPT > CKPT conversion to 'fix' those files. but that is more of a suggestion than a complaint.

Again, thanks for the wonderful tool N00mkrad!

n00mkrad commented 1 year ago

Any exampels of incompatible ckpt files?

PhSMu commented 1 year ago

sorry, forgot to check the thread for replies: At the top of my head, I remember converting at least Hentai Diffusion 21, which is known not to work with the program, and a bunch of OrangeMixes I got from the HuggingFace: https://huggingface.co/WarriorMama777/OrangeMixs . Other than that, I'm not really sure but maybe Grapefruit-Lemon and Anything V4.5.

PhSMu commented 1 year ago

oh, and on the first comment, this is basically what I mean by still getting a weird "0 memory bytes left" error when trying to generate some higher resolution images: Capturar as you can see, simple math dictates I should have enough memory for that: 4.78 + 0.4 = 5.18 < 6.00, so I have no idea of what exactly is happening here.

n00mkrad commented 1 year ago

What resolution did you try to generate

PhSMu commented 1 year ago

512x640, in this case, while on 1.7.2 I can do 640x960 or even greater like 960x960 or 1280x640