LTX-2 - Brought back anime tiddies in 1 hour 16 minutes at home on consumer GPU!
We all know the potential of a model reveals itself only when LoRAs are trained. This is my fourth LTX-2 LoRA after Hydraulic press, Inflate it and Cakeify. The LoRA brings the missing part of the human beauty (however, cartoon human, in this case). The LoRA has been trained fully on some dozens of synthetic big breasted Wan2.2 and the newer Kandinsky 5.0 generations - no real humans involved. 95% of the dataset was anime style, if you need realism, you can simply try the methodology on your own dataset.
The basic prompt is A cute anime girl pulling off her shirt, she is smiling kindly while touching her gigantic soft chest..
To accelerate the convergence and break the theoresized RL human body alignment (read more in my inflate lora description), TREAD, CREPA and FFN unfreezing were used. CREPA is a crucial component - read its paper https://arxiv.org/pdf/2506.09229.
Because the CREPA - representational alignment - technique was used, it's not recommended to use it in text2video mode. The picture is delicious, but sadly washed out because of CREPA - the cost of overcoming RL rapidly. :(
Limitations: The breasts may be too big for some of you and realism is out of the scope. Thankfully, LTX-2's quality is that warrants against the bad use :)
The training took slightly over an hour, literally (On a 5090, 1008 steps). This is a massive counterpoint to people saying against LTX-2 trainability on foreign concepts.
CREPA curve
The SimpleTuner training and dataset configs are under config.json and ltx2-multiresolution-t2v.json respectively.
If you think this cartoon LoRA is too problematic, please write in the Discussions page. Needless to say, this is only for adults!
LTX-2 generations are quick, so I won't attach illustrations here.
