ModelsLab/OC LoRA Tests - Kie Rose V1

Blind attempt at training a character with medium-level details with AnimagineXL-ObservationsTrigger word is a hard requirement otherwise the LoRA will not workRecommended Strength:1Clothing is fully removable and easily swappable but prefers long and descriptive promptsuse best_quality,masterpiece,modern for ease of nsfw according to AnimagineXL documentationStyle keywords are usable but I'm not familiar with all of the related keywords with AnimagineXLOverfittingAnime Screencap artstyle is bleeding over into the environment details causing generation images to have overly flat backgroundsmay require words such as 'head_out_of_frame' and 'from_behind' in negativeThere is still some level of style bleeding but I'm not completely familiar with how Animagine handles style keywordsStrong preference towards cowboy shotssome overfitting on face coloringUndertrainedWaist ribbon may or may not appearDoesn't seem to understand asymmetric so details can be flippedBelt color is undertrained but is occasionally correctBase outfit seems to lose detail when using other artstylesAugmentations:Img2Img ControlNet augmentation with Flat style lora and Add Detail LoRASwimsuit and underwear imagesNotable Differences from SD1.5Small rank compared to SD1.5 but SDXL has larger size overallremoved Style captioning to reduce overfitting of styleremoved Manual Pose regularization dataSingle Repeats folder for most of the data