Training A Character LoRA

hello this is Timothy from Scenario and I'm very excited to show you our new character training with sdxl and luras so the first thing I did is generate a character that I like on a signature Laura blend available on scenario I've developed a prompt that has given me good results so I've copied the image to my clipboard and I'm going to head over to chat GPT and upload the output we like and ask it to describe precisely the character in the image once we have the description we will ask a followup so that it gives us a concise 30w prompt we'll then take that prompt back to our generator and add add to the original prompt producing images from which we will gather a data set as a general rule for the settings I'm sharing we only need 8 to 15 images of a character but the possibilities are endless and we would love to see what you discover in your own experiments and creative Endeavors exploration and experimentation are the key we'll download what we want from the generated set showing a consistent looking character the settings we will use which are specific to this character workflow are 1,575 training steps which is calculated by the number of images in the data set times 175 the unet learning rate will be set at 65 and the text encoder training ratio at .2 and the text encoder learning rate is set at 1 E5 so let's put it in the oven now we can easily prompt our character into any scene we like here she is in front of a city skyline a forest and a desert with an extra smiling prompt stay tuned for more guides on our amazing features like blending multiple auras and even more consistency and control with reference image and control net image prompting combined

Channel: Scenario

Views: 16,951

Keywords:

Id: XHlg2xWJAes

Length: 1min 36sec (96 seconds)

Published: Fri Dec 22 2023