I've been benchmarking NightCafe's text-to-image for specific "text effect" prompts. Goal: clean, high-contrast typography you could actually use in a design. Most outputs are garbage.
My best result so far, using the `Coherent` model:
**Prompt:** `"devops" in glossy 3D letters, extruded, neon cyan outline, dark matte background, studio lighting, clean render, high detail`
```json
{
"model": "coherent",
"steps": 70,
"cfg_scale": 10,
"style": "enhance"
}
```
Key findings:
* `Coherent` handles complex typography terms ("extruded", "3D letters") better than `Stable`.
* High `cfg_scale` (10-12) is critical for prompt adherence on text.
* The word itself matters. Short, bold terms like "devops" or "k8s" work. Longer phrases fail.
Post your successful prompts and configs. I want to see:
* Your exact prompt.
* Model and key settings used.
* What made it work (or not).
Benchmarks or bust.