Working as a drop in for DaiSaWa Workflow
Dropped this in the Daisawa LTX2.5 Director workflow, Ran solid for 1st pass , on the refining 2nd pass melted the video entirely, swapped back to Euler-Beta 57 on the refine pass and it seems like its happy to cruise along, I know some DMD lora like to run at 50-75% should we just be dropping this one in at 1 for both video and audio? got really bad ghosting, but the audio was a SUBSTANTIAL improvement. this was with the distil lora at like 0.55 on Daisawa's NF4 model.
No it should be 1.0 both passes on all weights, it doesn't tend to do the destructive resampling of the distilled lora on upscale. I'm even running upscale sigmas: 0.92, 0.909375, 0.725, 0.421875, 0.0 - 5 steps with very high resample area and it usually adds details instead of random anatomy.
I think a hook is missing somewhere in the code, there is probably an event or hook that needs happen on your node to reply/stream the progress of the generation? Unless i am smoking my socks (AGAIN!)