This is meant as a companion to my Pulp Cover Art LoRA, which does a really good job of generating images in the style of mid century pulp magazine cover artwork without text. This LoRA can be used to add the text back in (if used with a positive weight), or to help further prevent text from being generated on more problematic prompts.

This was trained on mostly the same dataset as my Pulp Cover Art LoRA (500 images of artwork that was used on the cover of a pulp magazine), plus scans of the magazine covers that used the same artwork. (500 without text, 500 with text).

This was trained using "sliders".

Remember that positive weights will lead to a greater likelihood of text and titles being generated, negative weights will lead to a lower likelihood of text being generated.

In general, you will most likely start seeing text showing up on pulp-style artwork at around 0.6. In non-pulp-styles, it struggles a little bit more, and will tend to put text into places where it makes more sense, like on signs and background objects, and occasionally as tattoos on people, but it can generate text and titles on other style images with a high enough weight.

If you are using this to generate text, remember that this is still Stable Diffusion, so you are going to have a really hard time if you are expecting it to generate legible text. This mostly generates stable diffusion gibberish.