Description
👋 howdy, team; first thanks for the great app!
👻 After upgrading from v1.7.4 to v2.5.3 (also v2.5.1), I'm encountering issues on "Text to image" on non-default models ( stabilityai_stable-diffusion-2-1, v2-1_768-ema-pruned ) but not models ( Default_SDB_0.1 (default), Samaritan_3d_Cartoon_v1_s2 ) where the outputted image is always (blank, empty, black).
🙋♀️ I would expect this model to continue working agnostic to app version/upgrade. I also believe this is the main recommended version and would expect it to work out of the box.
🔬 Model v2-1_768-ema-pruned succeeds for me on v1.7.4 but not v2.5.x. My Mac is M2 Pro 32GB on Sequoia 15.3.1. To troubleshoot, I have ( uninstalled+reinstalled app, restarted computer, tested with all other apps closed, ensured no coinciding CPU/RAM strain ).
What appears to be coinciding non-sequential log errors during issue:
stable_diffusion/schedulers/karras_scheduler.py:129: RuntimeWarning: divide by zero encountered in log
stable_diffusion/stable_diffusion.py:565: RuntimeWarning: invalid value encountered in cast
Errors appear to slightly shift but symptoms remain when toggling "Advanced Options">Diffusion>Sampler from "karras" to anything else.
Examples on defaults with prompt "octopus" (with logs footnoted if erred as generated from fresh app start-up each time)
- (A) successful v1.7.4 with model
v2-1_768-ema-prunedfrom file manually downloaded from here months ago -
- (B) unsuccessful v2.5.3 with model
v2-1_768-ema-prunedfrom same save location as working just above - (C) unsuccessful v2.5.3 with model
stabilityai_stable-diffusion-2-1as installed today under "import models">"import from hugging face" - (D) successful v2.5.3 with model
Default_SDB_0.1as installed today - (E) successful v2.5.3 with model
Samaritan_3d_Cartoon_v1_s2as installed today under "Import Models">"Availble for download" -
