v4リリース
休養戻り。
瞳の潰れを対策。
チアリーディングにおける開脚ジャンプ。
チアリーディングでのジャンプ自体はタグ複合としてあるもののジャンプ位置頂点辺りにおけるピッとした足伸ばしがそうそう出ないため作成
ぶっちゃけcontrol netの方が安定感はいい
■注意点
瞳の潰れ対策に作り方を変更。
ちょっとやりすぎたせいで開脚ジャンプの足の角度、特に水平一直線出しの発生率が少し安定してない。
足の角度が少し下がりがち。
タグの構成を少し変えた。
■各推奨強度 学習強め
他LoRA使わない場合 推奨強度 0.3-
他LoRA使う場合 推奨強度 0.5-
全体映す場合、横長推奨
Adetelarは顔だけで可。
v4 Release
Returned after a short break.
Updated the training approach to address eye collapse issues.
This LoRA focuses on the cheerleading straddle jump.
Although cheerleading jumps already exist as combinations of tags, it is surprisingly difficult to consistently generate the clean, sharply extended leg position at the apex of the jump.
This LoRA was created specifically to reproduce that moment.
To be honest, ControlNet offers better overall stability, but this LoRA is intended for prompt-based generation without pose control.
Notes / Known Issues
The training method was modified to counter eye collapse.
Due to slightly overcorrecting this adjustment, the leg angle stability of the straddle jump is somewhat inconsistent.
In particular, the occurrence rate of perfectly horizontal, straight-line leg extension is not fully stable.
Legs tend to angle slightly downward.
The tag structure has been slightly revised compared to previous versions.
Recommended Strength
(Training remains relatively strong, same as previous versions.)
Without other LoRAs:
0.3+With other LoRAs:
0.5+
Usage Tips
For full-body compositions, landscape (wide) aspect ratios are recommended.
ADetailer is only required for the face.