Beauty of rain [Wan 1.3B]

Trigger Word: b3@ut1f0ll_r@in

Model: Wan 2.1 t2i 1.3B

All examples are generated with 1.0 LoRA strength.

For inference used Kijai's workflows

Training Details

This is my first attempt to teach lora for Wan 1.3B such a specific movement as painted rain. The dataset included 11 videos from the anime Kotonoha no Niwa (The Garden of Words) with lengths of 16, 30, 60, 72 and 81 frames. Diffusion-pipe was used for training. The training was conducted simultaneously at resolutions 256x456 and 480x854 with num_repeats=1. I used a slightly higher learning rate and as the best model I chose the 54 epoch with lr=8e-05 with LoRA rank of 32.