ModelsLab/OC LoRA Tests - Elrisv2 Loopback

This is a rather simple character to only test how LoRA training affects headwear, neckwear and clustered objects. Recommended Strength:0.8 ~ 1 for general usage0.5 or use fewer keywords if you want to change the bikini typeDataset:Original base image at 4096 x 11040 resolutionimage cropped into multiple smaller images for more detailused faceless/face censor augmentationStyle Augmentation for reducing style bleedLoopback training to help with outfit swappingObservations:Necklace and number pattern design was earliest than expectedGetting the flower next to the frills was harder than initially expected and doesn't generate all of the timeFace trains quicklyNotable ChangesTrained with prodigy optimizerDoesn't require a trigger wordTrigger word for Elris is included along with captions for numberbikini and alternativeoutfit exist. These keywords don't seem to have much of an impact but numberbikini can help with the success rate of the flower appearingLoopback images includedThis is mostly to help with outfit removability. Added multiple outfits + nude imagesRequires specific negatives such as head_out_of_frame, faceless, censored, identity_censor,face_censor in negative promptIssues/Difficulty:Posing is stiff and/or requires prompt emphasisHaving separate wings is difficultLoRA likes to place a lot of extra molesStyle Bleeding but can be mitigated with artist keywordsFace size is overtrained