It's very difficult to train gymnastic movements while maintaining accuracy and variety. The basic model doesn't contain gymnastic data, so I had to increase the grid size and train only the important layers. It also increases the text encoder, which provides better understanding. It's possible to compress LoRa to a smaller size and it will give excellent results, but the reverse cannot be done.

Just place the trigger, then add what you want to change or remove , I recommend using LCM