Closed catboxanon closed 1 year ago
Ah, good catch. The hr weights weren't defaulting to their low-res equivalents correctly.
2bd6608ca445cf7fd0384667eb9e8f35544c48cb extracts the weight setup routine to its own easily-testable function, corrects the behavior, and adds tests.
Currently, if a LoRA is entered in the prompt as something like
<lora:name:0.5>
, and highres fix is enabled, if you enable plotting the weights it can be seen it jumps back to1
in the highres fix pass. I've had to workaround this by entering<lora:name:0.5:hr=0.5>
.This is unrelated to #2 as I'm only entering the LoRA in the main prompt, and not the separate highres fix prompt.
I also don't think tests currently account for this which may explain how it got overlooked.