Why can't ANY model spell 'WHAM' without inventing a fifth letter?

by @elenamartinez_films in c/pop-art ·

Genuine rant. I have now generated 60+ comic bursts and not ONCE has a model correctly spelled a 4-letter onomatopoeia on the first try. 'WHAM' becomes 'WHAAM' (ironic, given Lichtenstein) or 'WHMA' or 'VVHAM'. 'POW' is somehow a coin flip despite being THREE letters.

I KNOW the answer is 'add text in post'. I do that now. But I'm baffled that in 2026, with everything these models can do, four capital letters in a speech bubble is still beyond them. They can render photorealistic water but not spell 'BAM'.

What is it about text specifically? Is it ever going to be reliably solved or is compositing type forever?

536 points · 6 comments

Comments

  • @ravimuellerlab · · 96 points

    The Lichtenstein irony of 'WHAM' becoming 'WHAAM' is too perfect. The model accidentally recreated an actual famous artwork's title through sheer incompetence. Pop art eating itself again.

  • @victorcosta_studio · · 99 points

    Text is hard because the models learned to render the STATISTICAL TEXTURE of text (letter-like shapes in plausible arrangements) without a symbolic understanding that 'WHAM' is a fixed sequence of specific glyphs. It paints the IMPRESSION of a word, not the word. That's why you get 'WHAAM', it's a plausible-looking word-texture.

  • @victorchen_vfx3063 · · 100 points

    'it paints the impression of a word, not the word' is exactly it. The model has no concept that letters are discrete symbols that must be correct. To it, text is just a visual pattern like brick or foliage. It's rendering text-flavored noise. Hence the fifth letter.

  • @yukimuellercreative4793 · · 52 points

    It IS getting better for common short words, fwiw. 'POW' and 'BAM' land for me maybe 40% now vs ~10% a year ago. Still not reliable enough to trust, so, composite. But the trend line is up.