RuneXX

/

LTX-2.3-Workflows

like 104

Actually tried something like that for LTX-2.0 and had something working quite well when LTX-2.3 dropped ;-)
Will try finalize it and see if i can make it work well with ltx.2.3

Dunno if this is related, but any plan to wire up LTX-2.3 22B IC-LoRA Union Control? I specifically need it for character consistency in I2V, I’m trying to do it myself with an assist from gemini, but there’s something I’m missing. At this rate I may just train a character lora, but I’m reluctant.. creating the dataset is such a chore, and these LTX2 loras need video clip data. Mine are all over the place because of inconsistency.

I know that may be asking a lot but these workflows work very well for me, and are a huge leap forward from wan2.2

It will unfortunately help little with character consistency . Its solely meant to transfer over motion or structure (support pose, depth and canny "controlnet" inputs).
So the "consistency" only goes as far as mimicking the input driving video to transfer over the pose or structure of that to LTX as a guidance.

For high character consistency a lora is the way yes (guessing you mean face and looks of the person).
Its the number 1 challenge with AI (both image and video) to have character consistency. Its no easy fix.

Training a LTX lora isnt crazy complicated, but its a bit of a catch-22 since you need consistent character videos to train the lora with .
That being said, LTX-2.3 seems to do pretty well if the first input image is a close up of the person.

IC Lora Union is more for something like this (transfer over pose for example) :

Ah youre right, it’s only canny, depth, and pose.. welp I guess it’s time to start curating my dataset

how about ic lora union + first frame by qwen edit/klein etc.?

how about ic lora union + first frame by qwen edit/klein etc.?

yes that what I ended up doing ;-) works pretty well ...
And easier than any masking etc like Wan Animate