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

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

Trigger Word: OIIA_cat
Model: Wan 2.2 t2v A14B
All examples are generated with 1.0 strength.
For inference used tazs-example-workflows

Gallery

The gallery for the HIGH model contains examples of working only with lora HIGH (with the exception of one comparative example). But for the LOW model, all the examples are a combination of HIGH+LOW, since the LOW model itself has virtually no visual impact on the generation result, and showing examples using it is a bit pointless.

Dataset Details

I wanted to try to teach something from the concept category, and a meme with a cat and a funny song came to mind.

And after watching one such video, you find yourself in a bottomless rabbit hole of YouTube with these cats in different variations. Most of these oiia cats can be found in the form of cheap 3D graphics in various funny scenes or even in video game mods. Since Wan sometimes commits the sin of generating 3D graphics itself, I thought that such a dataset with cats should be well suited for lora's training. So I put together 20 different training videos and generated prompts using the web version of Qwen. I purposefully removed any mention of spining and detailed descriptions of the cat's appearance from the description. I also had to throw out too abstract videos like this from the dataset (although I consider it a masterpiece).

Training Details

Initially, I wanted to train the 5B version, but I ran into a problem. The cat wasn't spinning! Although I have chosen different learning parameters. It has not yet been established what this was related to.

So I started experimenting with the A14B in small steps. At the first stage, I took ONLY ONE video with the original cat on a green background and prompt:

OIIA_cat, medium shot. The background is a solid, vibrant green screen. The uniform green backdrop indicates this may be part of a video editing process or special effects setup.

I wanted to get the results as quickly as possible, so I immediately used the lighting version and, lo and behold, some small fragments of the video with the cat were spinning! Success) I don't care that there are a lot of these cats and the image itself doesn't match the training one, these are the side effects of lightening, which I didn't give a damn about at the time.

The training schedule for 20 videos looked atypical for the LOW version. In my mind, the LOW version should have been noisier.

According to the visualizations, retraining clearly started at epoch 100, so 200 epochs is too much.

I've read that the best results are obtained when the HIGH version has fewer epochs than the LOW one, but in my case, both graphs almost matched. So in both cases, I took 70 epoch.

Prompt details

Unfortunately, in most cases, if you use only the trigger word OIIA_cat, the cat itself was not always spinning in place. Therefore, it is necessary to explicitly state in the prompt that it is spinning. I got the most consistent results with this description:

The [object] froze for a few seconds, and then abruptly began to spin in place.

although this description was not in the training set. Moreover, the object may not always be a cat. You can use a car or a telephone or even a stiletto shoe, they start spinning. Although sometimes there are artifacts in the form of suddenly appeared oiia cats :)

Version comparison

It was interesting to observe how the prediction changes if you use only the HIGH version. Everything looks very realistic, except for the spinning of the cat. The LOW version in its pure form has almost no effect on prediction, sometimes it can add a few grainy artifacts, but I do not recommend using it alone. The funniest visualizations that approximate the original meme appear when using HIGH+LOW. There are obvious grainy artifacts in places, but if you go through the seed long enough, you can get rid of them. But I like the effect itself.

I do not recommend using this lora with Lighting lora as it greatly distorts the result. Unless, of course, you want to get something even crazier and more broken than the original oiia cat meme.

Сonclusion

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