Tip: Placing trigger words together in parentheses improves consistency e.g. (urlilslt, speech bubble)
Tigger Words: urlilslt, speech bubble
Weight/Strength: 0.9-1; 1-1.3 when using other LoRA models (higher strength will improve text but may start frying image overall)
Prompting for specific text or including “text”, “writing”, etc in the negative or positive prompts can affect generation negatively. The LoRA was trained without prompting for text to not clash with Illustrious’s base knowledge.
A concept LoRA for generating a comic book style speech bubble of your character saying, "I'm your little slut". Only requires the two trigger words. Not very compatible with other text.
This is a hyper specific trial concept, so results are not 100% and can be influenced by other LoRAs, however the results are pretty good from my testing. Let me know if there are other super specific text you'd like to see as a LoRA!
Have fun, Degenerates! 😍
Description
V1 - Read model description for usage tips
FAQ
Comments (8)
Is PonyXXX-BJ1 used in some exacmples an embedding? I can't seem to find it.
Yes I rename my embeds for organisational reasons but check out:
https://civitai.com/articles/3776/list-of-models-signed-by-hentaidigitalart-pdxl-sd15
@Void91 has a ton of embeds for Pony/Illustrious and is one of the few creators actively making them. They have a discord too!
@degeneratemomofan Gotcha, and thanks for the reccomendation.
Sublime 😍😍😍
How do you even train a lora on a specific string of words?
LOL not super effectively but this is my attempt.
TL;DR: Don’t tag the text itself, tag the speech bubble and everything else.
For a more detailed explanation, the dataset I used had a bunch of various images with an inserted text bubble and the exact text. Then I refrained from tagging the text itself and only used the trigger word + “speech bubble”, then tagged everything else. This way the model learns “speech bubble” and exactly what is in the speech bubble without being influenced by Illustrious’s base knowledge. Attempting to use the tags “text” or trying to prompt the words themselves can actually mess up the generation because of Illustrious’s base knowledge of text.
@degeneratemomofan Yeah, I figured it was something like that.
I do find it a bit ironic though that an entirely alphabetical/numerical program has trouble generating words. Ironic in the funny way.
@WorstAirtist Truuue haha I've never thought of that. Illustrious does actually has a surprisingly good understanding of some text despite not specifically being trained for it. It might be more beneficial to tag the text because it still seems to be recognized as text anyways, but I wasn't sure.