These loras are intended to reproduce people blowing up large heart balloons. There is a relative scarcity of material depicting people blowing these, and most of the stuff is paywalled, which means a big lack of data available for these trainings, so I had to be very creative in my approach. More data should make it better.This lora is still not 100% accurate, but this is my best attempt so far before migrating to WAN 2.2. Prompts should be like these: "&lt;subject&gt; is blowing a giant heart-shaped &lt;color&gt; balloon with a long neck. He is blowing into the balloon. He is pressing his lips up against the balloon's opening. One hand is supporting the balloon by pressing it down to prevent air from escaping" You can specify the balloon's colour and maybe shape, but it is not guaranteed to work since the data pool was tiny. It will produce some funky shapes depending on the seed. Resolution-wise, it works best at 480p, but will deliver good results at larger resolutions. It will work with Wan 2.2 too, as long as you use it in the low noise stage. You can use lora strength of 0.5 -1.0. All the samples are generated using Phr00t's Wan 2.2 Rapid All-in-One checkpoint.

Heart balloon blowing - Wan 2.1 14B T2V

Initial release, will be updated sometime if I can gather more data

Files with SHA256: 72a7b8e44b3072941823ea93e5116e18382313fd63208ac13d7a4bee4585d0c3