Why nobody talks about how AI image generators mess up text in signs

I spent like 4 hours last week trying to get Midjourney to make a realistic storefront with a sign that said 'OPEN'. Every single time the letters were garbled or scrambled. I thought I was doing something wrong with my prompts. Then a buddy in a Discord server pointed out that most AI models just don't handle text well because they see letters as shapes, not words. That moment made me realize I had been fighting the tool instead of working with its strengths. So I switched to making signs with Canva and compositing them in Photoshop after. Has anyone else run into this weird limit with AI art tools and how do you get around it?

2 comments

2 Comments

luna8914d ago

Midjourney v6 actually handles text way better than v5. You gotta use the - style raw flag and put quotes around the text in your prompt. I tested it with "BAKERY" on a window front last week and it came out readable. Not perfect but way less garbled than before. Their newer models fixed this somewhat.

parker_park84d ago

haha yeah totally! i had the same thing happen when i tried putting "COFFEE" on a sign. it came out way cleaner than v5 ever did. the quotes thing is huge - i didnt realize you had to do that at first and was getting total garbage text. the raw flag helps too, i noticed it makes the ai focus more on the actual letters instead of trying to make them all swirly and artistic. i actually got like 4 out of 5 words readable on a storefront i did last week. still not perfect but honestly for ai art its kinda impressive. i think the newer training data just has a lot more examples of real signs and text from photos.