Teaching Stable Diffsuion to spell, one LoRA at a time!
Harrlogos is an SDXL LoRA trained on my own art and graphic design work, unlocking the ability to generate custom text.
It does not require any control net, any img2img, or characters from fonts on the canvas.
All you need is the LoRA, and you're ready to go!
The model comes with a plethora of styles, modifiers, and extra content built in, but that's only where it begins.
One of the greatest aspects of the model is how well and how seemlessly it works with almost any checkpoint, other LoRA, or even things like Stable Video Diffusion!
Harrlogos works with individual terms, separated by commas, in a specific order. You do not need all of these in every prompt, but this is the general format:
YOURTEXT text logo
Text Color: blue, teal, gold, rainbow, red, orange, white, cyan, purple, green, yellow, grey, silver, black
Accent Color
Background color
Style Modifiers: dripping, colorful, graffiti, tattoo, anime, pixel art, 8-bit, 16-bit, 32-bit, metal, metallic, spikey, stone, splattered, comic book, 80s, neon, 3D
Accent Modifiers: smoke, fire, flames, tentacles, hell, glow, horns, wings, halo, roots, embossed, blood, digital, ice, frozen, japanese, chrome, pastel, robotic, hearts, cute, egyptian, viking
Additional Content: cat, sword, owl, cat ears, cthulhu, sun, roses, clouds, space, stars, skeletons, demons, fog, trees, moon, skulls, bones, planet, earth, cherry blossom, pentagram, crosses, lightning, bolts, crown, circle, moth
Example of a prompt with the resulting image:
HarroweD text logo, white, grey, red, spikey, splattered, dripping, blood, hell, crown
Harrlogos is trained on a 1024x1024 data set, so it works best at resolutions: 1024x1024, 1024x768, 768x1024, 768x768
You will not get the correct spelling of every word, every time. Some text may take multiple tries to get right
The model works best with single words, multiple words are very achievable but it may be more difficult
Try weighting the text in the prompt, ex: ("Harrowed":1.5) text logo, red, black....
There's two schools of thought in achieving the perfect Harrlogo - A. Brute force spam through seeds B. Find a good starting point and lock the seed in, then slightly asjust the other paramters up or down, which will create slight changes to the image.
With a fixed seed, raising or lowering the CFG, Steps, or even resolution will modify the image, potentially fixing things or correcting spelling.
In the negative prompt, consider things like: Double letters, repeating letters, malformed letters, and more than X letters. If your renders keep coming out with a specific double letter, also try putting that in the negatives.
This is a work in progress!