Guernika

/

CoreMLStableDiffusion

like 65

Follow

Guernika 7

Also !
Can you pop this kind of paint/annotation menu on MacOS ?
I think its nice to guide inpaintings, instead of opening up other software

Hey @ZProphete I will take a look at that bug, I think I got it too at some point.

I will check again about adding conditioning to the shortcuts, I did try but I believe there were some bugs on macOS side, maybe it got better on 14.1.

About the inpainting, I do want to improve the painting but using the same as iOS is not possible, PencilKit is not available in macOS :/

I will also check IP-adapter, not sure if I have already, I was planning on checking out Blip diffusion, I thought it was similar to that reference-only pipeline but using an extra module, I will check those two and see if I can bring some to Guernika.

There is a new update waiting for apple that will add support for LCM models which are also very nice.

Sounds great, Thanks you for your work :)
The New Update is veeeeeery interesting, the LCM is quite FAST !
A few question about this :

• Where has the negative prompt gone, is it by the nature of LCM ?
• Will we be able to make custom LCMs in the converter ? I'm confused with the official description (LCMs can be distilled from any pre-trained Stable Diffusion (SD) in only 4,000 training steps (~32 A100 GPU Hours)
• This models removed the cap on 1024px, what about the older pre-update models ? (I could never make higher custom res anyway, my computer struggles and eats a big chunk of swap on 32g M1 Max)

May I suggest few more minimal stuffs :

• QuickLook preview API onto images would be nice, I tend to press spacebar while I'm navigating in the collections to see closer :D

• Is it possible to take inspiration from a certain detailer extension and return the positions of faces as a shortcuts action ? there is no cross-platform Shortcuts solution anywhere.
  (Although, I previewed YOLOv3 from Apple ML dev page and YOLOv8 from "john-rocky/CoreML-Models", they highlights full bodies AND faces as "person" so this will probably fail as is ?)

• "Share", "Add to Photo" drops the image in the library, any chance to drop it into an album ?

• LCM has a different architecture than regular SD models, it does not support negative prompt at the moment, that's also why ControlNet does not work either.

• You should be able to convert future LCM models yes, but they would have to be ready for that, you can't convert current SD models into LCM models without the extra training.

• What do you mean it removed the cap? the base resolution of this model is 768x768, older models have a base resolution of 512x512, variable size by default set the maximum resolution to be double of the default one.

• I will be implementing this QuickLook.

• What do you mean with this? what would be the output of this action?

• What do you mean with this too? Adding an external image into a collection?

I also wanted to ask, do you have any other suggestions/requests?
I was thinking on a major overhaul of the app, maybe have a list of jobs instead of the current Create tab, I was also thinking on redesigning the sidebar and collections, are you actually using collections? is it worth it to have it there? Does it make sense to keep the Models/Conditioning tabs in the sidebar or maybe have a joint Model library?

Hello,
Thanks you for the educational infos

• Well, what I meant is, my converted variable size models have maximum of 1024px, when I used the LCM one it was beyond 1024px, I was surprised.

• So, the goal to detect faces and return a box position X,Y, Width & Height; It would be used to enhance this little area only, bad/far faces like so : Txt2img > Detect face > Crop > img2img > Overlay new face.
  (Crop Action doesn't accept manual cropping, only numbers, it bothers me a little to open an image editor for this; Segment Anything or YOLO because AI is trendy, but any manual selection action would work, lol!)

• Nevermind, I wanted to sync some images from Guernika iPadOS and macOS using Photo, for now it saves in the root photo library, it could be great inside "Guernika Album" folder.
  or just put "Guernika_" in the filename so it can be sorted via "Smart Album"
  
  an iCloud ⛈ collection from Guernika is probably a no-go, people could think you collect images ^^
  Sorry I'm being very picky.

-You can probably have both,
the current Create tab has the "friendly" Mac-users crave for, and Job list for advanced usage ?

• Since you are going for QuickLook api, could showing smaller JPG(?) thumbnails in the collections benefit the memory ?
  I do use Collections; for organization, saving prompts, stuffs made with Shortcuts, and because it's easy to trash a "Test generation" folder from time to time.
  If you need this space put it in a tab, I'd arrange it like the Finder "group by" accordion, where a few are shown horizontally, and a button that collapses into one row.

But in this arrangement, if you show all images at once a solution to hide NSFW folders would be necessary.

• IMO You can merge those models tabs, once it's there I don't look at it very often, everything needed is shown in the main tab already, conditionings are auto sorted etc

I use Guernika on a large screen, I don't feel the necessity of reorganizing, but I can't speak for everybody !
It may lack a pinch of color here and there.
I don't know what you have in mind, so we will certainly give you feedbacks !

Hello !

• I'm failing at using LCM dreamshaper model on iPad M1 13", here is what I got from Console :
  

• I guess that iPad limitations prevented you from adding controlnet yet,
  I was wondering if you could at least add T2I adapter there, as the legend says it isn't heavier on the computations.

• Now that those LCM Loras are out, can it be merged and detected as LCM model ?

have a great day

Hello friends, do you have encountered this problem ?
"Suddenly" I can't use compression anymore on any models (1.5, XL, LCMs) it stops at the TextEncoder and shows this :

Hey @GuiyeC How are you doing with the UI, any teaser ? :)