This article aims to give you an overview of the process of creating your own style LoRAs and is best used in conjunction with the following articles:
Currently, Pencil offers the ability to create LoRAs with the base model Bria, and we hope to add support for other base models in the near future.
To get started, within your workspace go to Settings > AI Governance, where you will see an area called 'Custom Models'. This will show any and all LoRAs that have already been created, and has a button titled 'Train New Model':
Select 'Train New Model' and you will be presented with a series of inputs as well as an area to upload training assets. Don't forgot to reference How can I identify the best references for a Style LoRA?
Currently only Workspace Admins have the permission to Train custom models and make them available in the workspace. Custom models can also be deleted here by hovering over the model name and clicking the three dots for the context menu.
Using an example of a 'yarn art' style LoRA, I would complete the Model Properties as follows:
In addition I have uploaded eleven training images (a minimum of 10 images is required):
As outlined in How can I access and use my trained LoRA?, once trained your LoRA will show up alongside the other pre-existing AI models. With this in mind you may wish to choose one image to be the thumbnail and you can do this by hovering over the the image, clicking the three dots and selecting 'Set as thumbnail':
Once you have completed all the mandatory fields and uploaded a sufficient number of sample images, you can select 'Start Training'.
Note: Training can take three to four hours
Once your LoRA is trained and ready to use, you can go to Image Generation, select the model and begin prompting, and there are a few things to note.
In this example have selected my model 'Yarn Model'
Once selected, you will see the familiar Prompt boxes and if you expand the 'Generate Options' you will see a slider named 'Model Strength'. Broadly speaking, this allows you to control the amount to which your prompt leans on the LoRA vs the base model (Bria). In this example I am going to set it to 100 as I want my image to be totally influenced by my Yarn Model LoRA.
In your prompt it is important to reference your model and the keywords you defined. It is possible to have a prompt that, with strength at 100%, still fails to adhere to your style LoRA and draws from the base model due to a lack of specificity in the prompt itself.
In this example my prompt clearly specifies that I am seeking a yarn character and, with strength set to 100% you can see the results below. The generated image has clearly heavily referenced my LoRA and is a pleasing result that demonstrates an excellent combination of my Style LoRA and my prompt.
