I’ve been analyzing this issue from an observability perspective, and the problem of mangled text in generated images is a well-documented limitation of the underlying model architecture. While my expertise is in monitoring systems, not generative AI, the pattern is clear: these models are not trained for typography or textual accuracy.
Midjourney, and similar diffusion models, interpret text prompts as visual concepts, not as instructions to render legible characters. When you include a prompt like "a sign that says 'Open 24 Hours'," the model generates the visual *pattern* of text—shapes that resemble letters and grouped glyphs—but not coherent language. It's essentially producing a texture that statistically matches the idea of text.
From my testing, a few workarounds can mitigate the issue, though they are not perfect solutions.
1. Avoid direct text requests in the primary prompt. Instead, describe the *concept* of the text. For example, use "a cafe with a neon sign in the style of 80s retro typography" rather than specifying the exact words on the sign.
2. If specific text is absolutely required, you must use post-processing. Generate the image without the text, then use a separate tool like Photoshop, GIMP, or even a programmatic solution to overlay the correct text. This is analogous to adding custom metrics or tags in post-processing after log ingestion.
3. Experiment with the `--style raw` parameter or lower `--stylize` values. Sometimes this reduces the model's tendency to over-embellish textual elements into artistic shapes, but it is not a reliable fix.
There is no prompt engineering "secret" that will consistently produce correct text. The model's latent space does not contain a reliable mapping for orthography. This is a fundamental constraint of the training data and objective function, not a bug.
null