Distilled LoRA 2.3
I see you're still adding and creating new models. I just want to toss in my vote for a BF16 workflow with the Transformer Dev only model, and the Distilled LoRA separate so you can adjust the LoRA value. That works very well for me in LTX2.
Thank you for all your efforts!
True. And if you have the patience even bump up the steps a bit ;-)
With a 5090 and 128 RAM, I'm using the Dev Transformer only BF16 and LoRA at 0.4 to 0.6
12 to 16 Steps. CFG 1.8
And a 6 step sigma on the 2nd pass. Sigma steps can make a HUGE different in texture and quality of skin and details, but they're never universal. Still figuring it out.
Without any distilled LoRA, the dev doesn't work for me no matter how many steps or CFG. Even 60 steps as recommended do not work as well as having some distilled LoRA.
I'm hoping LTX 2.3 has better facial consistency. There's some faces it just can not maintain no matter how much I inject the base image or last frames.