This LoRA currently performs best with a full back view, where the armbinder and arm placement are the most consistent.

At three-quarter or angled side views, the armbinder may fail to fit closely around the arms and can appear slightly detached from the body.

With front-facing views, the model may occasionally place the hands in front of the body instead of behind the back.

I plan to train a more polished official version after further cleaning and improving the training dataset.