Image models still break long text. In my test on 14 November, Krea 2 Turbo broke a third of the words in a 52-word note. So I stopped asking the model to write the text.

Now the model draws the surface as a blank magenta sheet, and my command makeImage maps an HTML page onto it. It finds the sheet's corners, removes the perspective, follows the edges and any curve, and takes the light from the magenta itself, so shadows and creases land on the print. Every letter is exact, because it is a web page, not a guess.

It has limits. It keys one flat surface per run, and a soft sheet in someone's hands bends the text with it. I also put it up against Google's Nano Banana 2 on one hard prompt, one picture per model, so treat that as a single test.

The full post has:

  • the cover prompt recipe I use
  • a setup prompt that has Claude install ComfyUI and Krea 2 for you
  • a side-by-side with Nano Banana 2, and what each costs

Read the full post, with the setup prompt and the comparison: https://www.digitaldias.com/blog/2026-11-21-local-ai-blog-cover-images/