I believe I've found an issue with current hijack implementation: model_hijack.hijack(clip) ignores clip patches applied by LoRAs. This affects all parsers except comfy and comfy++.
You can verify this by changing strength_clip value in Load LoRA node, there will be no differences in output images.
I believe I've found an issue with current hijack implementation:
model_hijack.hijack(clip)
ignores clip patches applied by LoRAs. This affects all parsers exceptcomfy
andcomfy++
. You can verify this by changingstrength_clip
value in Load LoRA node, there will be no differences in output images.