Beauty of rain [Wan 2.2 TI2V-5B]

https://civitai.com/models/1747192/beauty-of-rain-wan-2122

Trigger Word: b3@ut1f0ll_r@in
Model: Wan 2.2 ti2i 5B
All examples are generated with 1.0 LoRA strength, CFG=5 and steps=30
For inference used Kijai's wanvideo_2_2_5B_I2V workflow with a small modification for t2v mode. You can find a more detailed description in article Hollow knight [Wan 2.2 TI2V-5B] in the section Workflow Details.

Training Details

The same dataset was used as for Beauty of rain [Wan 2.1 T2V-14B]. I'll say the obvious thing, but the 5B version is much better than 1.3B. From now on, I will no longer train loras for the 1.3B version, since 90% of the generations of 1.3B there are trash. The 5B version is much nicer in this regard, but it is also quite capricious towards prompts. As a result, 5B lora has a more realistic style than 14B. But I like these illuminated outlines of people against the background of the rain. I consider this a pleasant side effect, since there were no people in the training dataset at all, only raindrops and waterfalls.