'Tomboy' written on a shirt, 1girl wearing a shirt, living room, red hair, short s-76394925.png

Notices where this attachment appears

1. :back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War (zemichi@bae.st)'s status on Sunday, 03-Dec-2023 23:52:44 JST :back_from_gab: :musashi_berseker: :abigail_williams: :musashi: :benienma: :astolfo: :chloe: 🏆🏆 winner of the Holy Gab War and Holy Poast War
   Generating text in AI art
   
   :emergency: WARNING :emergency:
   This stuff is schizophrenic nonsense. There may be superstitions and placebos ahead.
   
   SDXL apparently can generate text, and because NAI V3 is based on that I decided to look into it to see how it is done. Right off the bat you should know that this is unreliable and it's way easier to just edit in your own text through different methods. This will cover generating text through a prompt.
   
   First, text generation in an image requires prose. Now I'm not good at prose so others can explain this better, but the general rule is that when writing prose for a prompt is that you write in positives and adding rather than negatives and subtracting. Like WITH instead of WITHOUT. And the prose should be simple. In text generation it considers the whole prompt for text so unrelated words can generate instead.
   
   Second, samplers matter in text generation. ancestral samplers should be used. In NovelAI V3 the ones that are effective are "Euler Ancestral, DPM++ 2S Ancestral, and DDIM" and the SMEA and DYN options should be off.
   
   (Theory: I took a minute to read up on these samplers. Ancestral samplers converge differently and cause more variation in images generated. I'm guessing this variation allows text to generate easier. Why? I don't fucking know.)
   
   Third, wording matters in prompt. I tried to read up on what people normally use to generate text and there are a few set ups:
   -Text "Text Value" written on
   -'text value' written on
   -saying "text value"
   -TEXT VALUE written on
   You get the idea. People use " and ' for text values, sometimes CAPITALS too. I got the best effect from quotation marks.
   
   Fourth, text gen becomes easier when there's a word or letters it recognizes. So simple words like "love" get recognized easily. Complex sentences will fail.
   
   Issues:
   Because I can only use NovelAI for testing I'm limited to what I can really test. In fact, I think NovelAI V3 may be more limiting for image text gen because of its anime nature. For NAI the rule you should follow is if the object is supposed to have text on it you can write prose to have it print on that element in the image. So a T-shirt, a sign, something that can have text on it.
   
   Here are a few sample images for text gen. I used those three samplers I've mentioned in these. :back_from_gab: #zemAI
   Text 'fuck you' written above 1girl, 1girl, medium hair, green hair, goat horns, s-1034245795.png
   Text 'fuck you' written above 1girl, 1girl, medium hair, green hair, goat horns, s-1199645115.png
   'Tomboy' written on a shirt, 1girl wearing a shirt, living room, red hair, short s-76394925.png
   'NEET' written on a shirt, 1girl wearing a shirt, bedroom, bed, messy hair, gree s-1117684108.png
   'DESU' written on a school swimsuit, 1girl, pool, long hair, low twintails, ligh s-108802295.png