LTX 2.3 DEV INT8 Convrot + Scaled V3 and more ....
I need an LTX 2.3 DEV INT8 ConvRot + Scaled V3 model, and I'd like to suggest one more idea.
Most people use the LTX 2.3 DEV model together with the Distill LoRA at weights such as 0.5, 0.6, or 0.7. However, in this setup, the INT8 acceleration benefit is lost because the LoRA is still applied at runtime, making inference noticeably slower.
Would it be possible to create a model like this?
LTX 2.3 DEV INT8 ConvRot + Scaled V3 + Distill 384-rank LoRA (0.5 weight)
I tried merging it myself. The Scaled V3 merge seems to work well, but after adding the Distill 384-rank LoRA (0.5 weight), I noticed some subtle artifacts or abnormal behavior.
This is the process I followed:
Merging by INT8 w8a8 node (without On the Fly)
BF16 DEV + distill 384rank (0.5) = BF16Model A
Merging by INT8 w8a8 node (with On the Fly + Convrot)
BF16Model A + ----> INT8Model A
combine block to made Sclaed INT 8
BF16 Model A (1,2,46,47) block only + INT8Model A = INT8Model B
INT8ModelB = Scaled INT8 Convrot + distill 0.5
is it right?