Modify the weights of the LoRA to achieve the desired result.
This was also trained with the expanded VAE located here. Please use this VAE when inferring with this LoRA.
Description
This is a brand new training using the original data set from 1.0, but resolution set at 768x768. Trained the original data set at 9,000 steps, then trained full body at another 2,100 steps. Standard portraits at 512x512 and 512x768 will generate a lot of close up shots. Note the resolutions below for more interesting full body shots. You will likely still have to inpaint the face at a distance, but it should be improved from version 1.0.
Please note that the 'trigger' word has changed for this version - it is now: "laurenboebert woman".
Sample pictures generated with the base SD 1.5 model.
Try with the following resolutions:
512x512
512x768
448x1024
512x1088
384x1216
Training Data:
Used bmaltais's implementation of kohya_ss LoRA training scripts.
Total Images: 15 total high resolution images: six full body, four head shots, five chest-up
Regularization Images: 300 total "sexy woman" regularization images from ProGamerGov's dataset at huggingface.
Trigger word: laurenboebert woman
Training Steps: 11,000
Unet Training Rate: 5e-05 (0.00005)
Text Encoder Training Rate: 5.0e-05 (0.00005)