Prompts in parentheses are optional.

V2.1 Update

V2.1 improves pose consistency. In most cases, you can now generate the target pose reliably without increasing the LoRA weight.

This version still shows some signs of overfitting and may conflict with other art style LoRAs or @artist prompts. Using many style LoRAs, especially at higher weights, may noticeably reduce image quality (softer or less detailed).

As a workaround, you can try generating the pose first with this LoRA enabled, then disable the LoRA and regenerate the image using the original result as a ControlNet Depth or Canny reference. This can help preserve the pose while allowing other styles to render more cleanly.

This is not intended to be the final version. V3 will be released after I find a better solution to the compatibility and overfitting issues described above.

V2

This LoRA's stability is somewhat limited, so if the pose does not come out correctly, try adjusting the LoRA weight and testing different strengths.

The LoRA currently struggles with front views (from_front) compositions, and results are much more reliable from_side, from_back, or three-quarter angles.

I may create a V3 in the future if I can collect a larger and better dataset.