A dedicated workflow for my current vpred model, unfourtunately hitting a Math special case, causing drastic difference between WebUIs and ComfyUI.
Visit my account page for Tech Report.
It is likely occurs when you switch objective variable (e.g. eps to vpred, vpred to RF) from a finetuned model. It will eventually fade off (null uncond matches zero uncond) when more GPU hours are baked in.
Description
I think no one are insane enough to train models / LoRAs with different objective.

