Bugs and improvement suggestions for Guernika app

#8
by D1ffus10n - opened

I bought the app recently and was playing around with it a bit. I also use Mochi Diffusion and Prompt2Image. I was hoping to benefit from the inpainting functionality using CoreML which is not implemented in the other two apps so far. Unfortunately the app still has a lot of bugs and quirks which will hopefully ironed out in future versions. Some of my observations:

  • Some models that work fine in other apps just crash Guernika on load
  • There is no way to see the generation data of an image from within the app (copying prompts and seeds should be possible easily)
  • The gallery is very confusing. I still have no idea what the purpose of those subfolders is in all images?!
  • In inpainting mode, if you want to change the base image, there is no such option available. It only works by changing tabs, going back, clearing and then choosing another image. The mask though stays even though the base image was changed
  • Also the mask option overlay window disappears randomly (same workaround necessary to get it back)
  • The full prompt data should be saved with the PNG
  • What is the Copy command intended to do in the context menu of an image
  • You cannot get an image from the gallery into the inpainting window
  • Adding external images (even when 512x512) to the inpainting window, unloads the selected model and/or crashes the app on generation
  • I had a two cases of memory leaks that crashed my whole machine
  • This is probably a CoreML limitation but inpainting seems to very rudimentary in comparison to WebUI.

And finally a question: What is the recommended inpainting model atm (1.5.or 2.1)? What are the differences?

Keep up the good work and I'm sure in the near future we will have a few really good SD apps that will come close to what is possible with WebUI or Invoke with the rocket speed of Apple Silicon Neural Engine!!!

Guernika org

Hey @D1ffus10n ! Thank you for your comments I'll address quickly the ones I can answer and I'll check more carefully the rest of them.

  • Can you tell me which models are you having problems with? It could also be caused by the compute units you chose when loading the model.
  • I know this and I will be working on improving reusing prompts and seeing the prompt used.
  • I believe the subfolders you are seeing are where the image variations are being safe, the idea is you may be working with multiple different images and you may be generating lots of variations for each of them, so you can see all the variations on this "subfolder".
  • I will take a look at this.
  • And this too.
  • The prompt data is saved with the PNG on its metadata, you can see it on Finder, but again, I will work on improving this on the app itself.
  • The copy command basically adds the image to the clipboard so you can paste it somewhere else, on another app for example.
  • Selecting Generate variations should do exactly this.
  • I will check this.
  • This will be hard to check but if you have a workflow that causes the crash please let me know, it could also be some memory leak on CoreML itself and for that the only solution would be waiting.
  • I believe we could get the same inpainting/outpainting as other apps, at the moment is a very simple "paint a mask" flow, I don't really know what is possible on WebUI but it should also be possible on Guernika, is there anything specific you would like to see?

About the inpainting model, I'm not really sure, I guess they would work differently depending on the prompts, you should use the one you are more comfortable with, maybe in Reddit there are better explanations on what the difference actually is.

Finally, thank you again for this detailed comment, I will keep working on this app and hopefully get close to those apps you mention.

Sign up or log in to comment