OIIA (Spinning Cat) [Wan 2.2 T2V-A14B]

https://civitai.com/models/2018677/oiia-spinning-cat

Trigger Word: OIIA_cat
Model: LTX-2 19B dev
For inference used tazs-example-workflows
The official training script was used for training.

About the LTX2 model

Well, it's happened! Thanks to the open source LTX-2, the era of silent films is over. Actors now have voices, and cats can meow ☺️

Compared to the Wan, the LTX-2 is much, much faster in both inference and lora training. Yes, it has some glitches and artifacts when generating videos, but who cares? Now you don't have to wait hours for the generation result. Everything happens incredibly quickly, in just a couple of minutes. I'm eagerly awaiting the next version of LTX 2.1 😎

Training

For training, I used the original code from the LTX2 authors, but I ran into some issues. A closer look at the repository reveals that the training script requires at least 32 GB of VRAM. Since I only have a video card with 24 GB of VRAM available, I tried running the training on that one first. I had to downsize the dataset to a microscopic 192p resolution, load dev-fp8 weights, and set the load-text-encoder-in-8bit flag. This only allowed me to run at rank 32. Increasing it to rank 64 resulted in an OOM error. This is despite the fact that I completely removed sound from the training...

After this, I should have switched to ai-toolkit, as it seems to have special optimizations for 24 GB video cards. But I couldn't find any confirmation that anyone had successfully trained lora on 24 GB using the dataset's appropriate resolution.

So, I decided to go all-in and temporarily got a video card with >24 GB VRAM 😈 This allowed me to train at 480p resolution with sound and a maximum video length of 73 frames. I left all the training parameters at their default settings.

Dataset

I had to slightly modify the original dataset. This is because I initially downloaded the videos without sound, and when I tried to find them again online, I couldn't... Some of the older videos weren't suitable because extraneous voices were clearly audible in the background of the cat's song. I also added several new videos where the classic cat song is slightly modified to resemble covers of famous songs. A total of 23 videos were used for training.

Conclusion

Overall, I like the way ltx2 handles. However, the video tends to fall apart when there's too much movement. If you turn off the sound, Wan 2.2 certainly wins in quality. But if you turn on the sound and watch all the cat videos, then

O-i-i-a-i-o, o-i-i-i-a-i
O-i-i-a-i-o, o-i-i-i-a-i
O-i-i-a-i-o, o-i-i-i-a-i
O-i-i-a-i-o, o-i-i-i-a-i

O-i-i-a-i-o, o-i-i-i-a-i
O-i-i-a-i-o, o-i-i-i-a-i
O-i-i-a-i-o, o-i-i-a-i-o, o-i-i-a-i-o, o-i-i-a-i-o
O-i-i-a-i-o, o-i-i-i-a-i, o-i-i-a-i-o, o-i-i-i

O-i-i-a-i-o, o-i-i-i-a-i, o-i-i-a-i-o, o-i-i-i, o-i-i-i
O-i-i-a-i-o, o-i-i-i-a-i, o-i-i-a-i-o, o-i-i-i, o-i-i-i
O-i-i-a-i-o, o-i-i-i-a-i, o-i-i-a-i-o, o-i-i-i, o-i-i-i
O-i-i-a-i-o, o-i-i-i-a-i, o-i-i-a-i-o, o-i-i-i, o-i-i-i
O-i-i-a-i-o, o-i-i-i-a, o-i-i-a-i-o, o-i-i-i, o-i-i-i
O-i-i-a-i-o, o-i-i-i-a, o-i-i-a-i-o, o-i-i-i, o-i-i-i
O-i-i-a-i-o, o-i-i-i-a, o-i-i-a-i-o, o-i-i-i, o-i-i-i
O-i-i-a-i-o, o-i-i-i-a, o-i-i-a-i-o, o-i-i-i, o-i-i-i

O-i-a-i-a, a-a
O-i-a-i-a, a-a
O-i-a-i-a, a-a
O-i-a-i-a
O-i-i-a-i-o, o-i-i-i-a-i
O-i-i-a-i-o, o-i-i-i-a-i

O-i-i, o-i-i, o-i-i, o-i-i-a-i-o
O-i-i, o-i-i, o-i-i, o-i-i
O-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o
O-i-i-a-i, o-i-i-a-i, o-i-i-a-i-o, i-i

O-i-i-a-i, o-i-i-a-i, o-i-i-a-i-o, i-i, i-i-i
O-i-i-a-i, o-i-i-a-i, o-i-i-a-i-o, i-i, i-i-i
O-i-i-a-i, o-i-i-a-i, o-i-i-a-i-o, i-i, i-i-i
O-i-i-a-i, o-i-i-a-i, o-i-i-a-i-o, i-i, i-i-i
O-i-i-a-i, o-i-i-a, o-i-i-a-i, o-i-i, i-i-i
O-i-i-a-i, o-i-i-a, o-i-i-a-i, o-i-i, i-i-i
O-i-i-a-i, o-i-i-a, o-i-i-a-i, o-i-i, i-i-i
O-i-i-a-i, o-i-i-a-i, o-i-i-a-i-o, o-i-i-i-a-i