LTX-2 Cakeify

Model description

LTX-2 Cakeify

We all know the potential of a model reveals itself only when LoRAs are trained. This is my third LTX-2 LoRA after Hydraulic press and Inflate it. It follows the quest of porting some the classic VFX LoRAs to LTX-2 (which has sound!).

This is the second attempt, the first one was with only the attention being tuned. Here, following inflate-it, in addition to CREPA and TREAD, I also unfroze FFN ("ff.net.0.proj", "ff.net.2"), however I compensated for its parameter growth by lowerting the overall LoRA rank and the size gain was not that large compared to H-press (830 vs 640 Mb, and 1 Gb on inflate). The LoRA was trained in 2 hours 41 minutes (1400 steps) on a single 5090, which I find is medium complexity, and the LoRA is pretty robust to various object types. If the object is stubborn, replace its name in the prompt with "cake" and it will be easier for the model to understand it.

The LoRA, in contrast to H-press or inflate-it, has been trained on pure, real videos of objects which have had turned out to be a cake after being halved.

The thing I love the most about LTX-2 is that you are able to set or generate any background music, be it a song or a greek style tune.

The SimpleTuner training and dataset configs are under config.json and ltx2-multiresolution-cakeify-t2v.json respectively.

Trigger words

You should use c4k3 cakeify it to trigger the image generation.

Also, turn on the sound or prompt or add your own!

Download model

Download them in the Files & versions tab.