Now that we have FLUX, it is possible to train Text/Typeface Loras!
Turns out you can train the model on the text only in a specific typeface style+text captions
As result, the model is able to generate texts, typefaces, logos, emblems, etc in a trained style of Y2K with a high precision. The future is bright
I'll be releasing more text/typeface Loras soon
Enjoy
Description
FAQ
Comments (8)
awesome work, can we train the model on other language typeface ?
Gonna find out soon. There's a chance that basic captions on other languages will be sufficient for Flux to pick it up
I did the tests and unfortunately it's not possible to train a language with just a Lora. The results weren't even close. There's a chance that perhaps a full fine-tune on different language could yield good results. Otherwise the only hope is training a hypothetical transliteration model. i.e. the model would generate text on desired language, but you'd have to prompt it with transliteration (basically, latin symbols).
@VadimFedenko i just need to know any guide that could lead me to something close. thx for your feedback
Good find
I’ve been trying to do this with replicate with varying success. I’ve been getting python to create images of random text in a specific font then have it attach a txt describing the image and a trigger word. I use that as the training data. Is that a good way to do it? I’d be keen to know how you set up your training data if you are open to sharing that?
yeah i basically used the same workflow. make sure the text in the dataset is easy to read and short (like 1-3 words), then tag the words in the images like "text: 'words' "
could you make one of these for illustrious?
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.