Can you reliably generate text on images yet?

I've been stress-testing Leonardo's text-to-image capabilities for a project where we needed to generate simple branded graphics with clear, legible text (think event banners, promo tiles). The goal was to see if it could save on design costs. The short answer: No, not reliably. It's a cost trap if you need precision.

Leonardo, like most diffusion models, treats text as a visual pattern to be approximated, not as semantic information to be rendered correctly. You can sometimes get a readable word or short phrase after dozens of generations, but it's a lottery.

Here's a concrete example from my tests:
* **Prompt:** `A rustic wooden sign that says "Welcome to the Cloud Cost Summit" in carved letters, photorealistic`
* **Result:** The sign looks great. The "text" on it might show garbled characters, merge words ("CloudCostSummit"), or have entirely wrong letters. Consistency across multiple assets is impossible.

If you're generating concepts where text ambiance matters more than accuracy (e.g., a blurry neon sign in the background), it's fine. But for any functional text, you must generate the image and then overlay the text in a separate tool like Canva or Photoshop. This adds a step and negates the supposed efficiency.

From a cost-optimization standpoint, using Leonardo for this specific job is like using a Lambda function for a long-running batch job—the wrong tool for the requirement, leading to wasted credits and time. You'll burn through tokens on re-rolls with no guaranteed ROI.

For now, the only reliable architecture is a two-step pipeline:
1. Generate your background/image asset in Leonardo.
2. Composite the required text onto it in a proper design application.

Until the model fundamentally understands typography and spelling as a discrete layer, this isn't a solved problem.