Scale-Bender — Tsunade · AoRA v2
Base model: AnimaYume v4
CHECK CHANGELOG >>>>
Most character LoRAs are trained on vibes. This one was trained on structure.
AoRA is my character LoRA workflow for the Scale-Bender project. It is not a new model architecture, but a dataset preparation, curation, captioning, and training methodology focused on character preservation.
This version focuses on identity and body-structure consistency, using morphology-aware captions and a controlled dataset split. The goal is to keep the character recognizable across poses, styles, compositions, and outfit variations.
Sakura stays Sakura.
Share your generations and feedback in the comments — every result helps improve the next AoRA version.
TRIGGER: sakura_aora, 1girl, body reference, official outfit, navel
This model not permits users to:
Sell images they generate
Use the model without crediting the creator
This model permits users to:
Run on other service
Run on Civitai
Description
Initial public release.
Changelog:
- First Sakura AoRA release.
- Trained with the AoRA v2 character workflow.
- Focused on character identity, body-structure consistency, and outfit stability.
- Uses morphology-aware dataset preparation with official / fanart / body-reference separation.
- Designed to keep Sakura recognizable across official outfit, alternate outfits, dynamic poses, and different compositions.
- Tested on portrait, full body, action pose, alternate outfit, and wide-shot prompts.
- Recommended for anime and semi-realistic anime checkpoints.
Known limitations:
- Clothing and scene tags are still mostly handled implicitly in this version.
- VLM-based semantic captioning for outfit, accessories, and scene context is planned for future AoRA versions.
- Some checkpoints may alter the face shape or outfit details more strongly than others.