Let's share LoRAs!

[kabachuha]

Давайте делиться лорами!

Here is my first one, thanks to kohya-ss/musubi-tuner#774.

This replicates the known "Squish" effect from one of the most popular Wan LoRAs.

Trained the I2V Pro LoRA on 256x256x73f fitting just into 24 Gb VRAM with 20 block swap and fp8, with ~6 s/it on 5090. Let's see how much can I increase the resolution/length in this regard. Despite the PR targeting the T2V model, it easily transfers to I2V because of not much difference between their architectures. Overall, this LoRA training took ~2-3 hours. The ComfyUI workflow is embedded in the mp4 video.

I'll be keeping my LoRAs here: https://huggingface.co/kabachuha/my-kandinsky5-video-loras

Kandinsky-Squish.mp4

---

Share yours!

[HapKoM]

Thank you for your contribution! Squish and cakeify video effects look amazing!

[kabachuha]

I noticed a potential problem, however, with the model's physics or logic. When I trained the model on hydraulic press crushes, it's visible that the objects start bending before the press hits them. The original dataset (OmniVFX/Crush) doesn't have this premature bending and when I trained Wan2.2 on it in its time, the hydraulic press crushes went as expected. I think I encountered such action bleeding in the vanilla Kandinsky model as well, need to gather some representative prompts.

Kandinsky_00041.mp4

[kabachuha]

Uploaded the proper Hydraulic press LoRA. This one was surprisingly hard to make. The post above used musubi-tuner and I couldn't get good results from it.

For proper training, I used the new CREPA technology consisting of semantic alignment between the pairwise video frames and features from a powerful teacher.

However, vanilla REPA/CREPA are hard to maintain because of the need to hold both VAE/DINOv2 into memory, absolutely destroying VRAM capacity.

Thanks to the newer LayerSync paper, released this autumn and SimpleTuner's maintainer bghira, the features were introduced not from an external encoder, but from the deeper layers of the network, eliminating the need for extra VRAM.

Here are some examples:

Kandinsky-Crush-En.mp4

Kandinsky-Crush-Ru.mp4

The LoRA is bilingual. (RU/EN)

If the Kandinsky team is reading this, I'd suggest exploring CREPA/LayerSync perspectives for model training / SFT 🙂