First attempt at an LTX-2 Character LoRA.
Trained with ai-toolkit @ 6000 steps.
Really impressed with the voice replication on this model.
Description
Switched to LTX2 base model.
FAQ
Comments (10)
Fanstastic, so authentic when it hits the animation style true. Do you have any tips to help guide it to get the "animated cartoon" look reliably? Any stronger terms to force the animated cartoon style out of it? The character and voice is alwasyy perfect, just still seems keen to give a realistic or 3d style.
Thanks for the great work!
try
"animated cartoon j3ss1c4 with long red hair is wearing a long pink sparkly dress with a slit up the side and purple gloves."
This seems to keep the authentic look for me. You may need to rotate to find a decent seed.
@Goon_69420 thanks. i'm using distilkled. i usede your prompt and it hits sometimes, i'll let you know if i find anuthing that helps steer the underlying ltx2 model to align more often.
@TheFunk which workflow are you using, and which distilled model?
@Goon_69420 I was gonna ask the same thing. At least for i2v, the style always shift back to realistic 3d character, using the mentioned prompt does not help unfortunately. I use ltx-2 phr00tmerge diffusion model, I believe it's base ltx-2 but with a truck load of Lora integrated at its core.
Finally, the women in LTX-2 have some curves
none curvier than Mrs Rabbit!
This is impressive! What did your training dataset looklike? I'm trying to train character lora with audio :)
36 video clips of 5s @ 24fps
any way block swap in ai-toolkit ? like in musubi_turbo