Intention

This LoRA was created to aid in the creation of letterforms / characters using the img2img process. The starting point was the question whether it is possible to create clean, typographic charaters using Stable Diffusion. The main focus of this LoRA is the creation of decorative / display fonts.

It is trained on a custom dataset on high contrast typographic imagery with flexibility in mind.

Base model for the training was v1-5-pruned.safetensors.

Use

This model is intended to be used in the img2img process. Generic letterforms (ideally white on black background) are used as an input. Playing with a noise fill for the letters or a blur can lead to interesting results, too.

To use the LoRA just include the following tokens in the construction of your prompt:

zhuj, typography

A starting point for experimentation with reasonable flexibility:

zhuj, typography, black background, monochrome, (((style description))) <lora:zhuj_typography:1.5>

Further refinement can be achieved using the Negative Prompt.

CFG scale of around 10.5 and denoising between .7 and .9 are a good. starting point.