This 16-vector embedding was trained using ADAMW with a constant scheduler (LR 0.00075, 350-step warmup); I configured a 12-image batch size with 25 steps per epoch which ran for 111 epochs (2775 steps). The training data consisted of 60 body photos (1x repeat; attention masked) and 80 cropped face closeups (3x repeat; attention masked), meaning each epoch ran through 300 images. The training dataset included a ton of pinup photos, so the outputs may trend risqué; I'd strongly recommend taking that into account with your negative prompts if you want to prevent NSFW generations. Training software utilized is the awesome OneTrainer (found here: https://github.com/Nerogar/OneTrainer).
This 16-vector embedding was trained using ADAMW with a constant scheduler (LR 0.00075, 350-step warmup); I configured a 12-image batch size with 25 steps per epoch which ran for 111 epochs (2775 steps). The training data consisted of 60 body photos (1x repeat; attention masked) and 80 cropped face closeups (3x repeat; attention masked), meaning each epoch ran through 300 images. The training dataset included a ton of pinup photos, so the outputs may trend risqué; I'd strongly recommend taking that into account with your negative prompts if you want to prevent NSFW generations. Training software utilized is the awesome OneTrainer (found here: https://github.com/Nerogar/OneTrainer).