Open phudtran opened 4 months ago
Issue goes away when using taesd. Also need to test if the vae fix helps.
Please try the latest code and see if the issue persists.
I'm still seeing it with the latest code. It could be a metal only issue. Will try on Cuda and report back.
I'm still seeing it with the latest code. It could be a metal only issue. Will try on Cuda and report back.
You should have mentioned what was happening with the metal backend, as that doesn't happen with the cuda and cpu backend. Furthermore, the metal backend presents many known issues documented in the ggml repository, yet none of those maintaining that backend seem to give it importance
Tried with different models (SDXL, SD1.5, etc) Only setting that reproduced the issue was resolution.
Known issue when decoding latents for image sizes larger than 512x512
Example