Sage Vermillion
@sage · March 15, 2026
Our moderation pipeline, what we automatically reject, and the human review loop behind reports. Some of the philosophy, more of the boring detail.
Introduction
Most prompts fail not because the model doesn't understand them, but because the prompt itself doesn't actually describe a single, coherent image. The fix is structural — start from the brief, not from a list of adjectives.
In this article we'll walk through the exact framework we use internally when reviewing model output: how to think about subject, composition, and render quality as three independent layers, and how to test changes to each in isolation.
Why it matters
Spending three minutes structuring a prompt almost always beats spending fifteen minutes regenerating until something looks acceptable. The cost isn't just tokens — it's the time you lose to ambiguity.
“A prompt is a contract. The clearer the contract, the fewer arguments you have with the artist on the other side.”
— Mira Aether, prompt design lead
The method
We break a prompt into four canonical pieces:
- Subject — who or what is in the frame, and what they're doing.
- Setting — where, when, what time of day, what light.
- Composition — camera, framing, lens, distance.
- Render — style, medium, post-processing language.
Each piece can be tuned independently. If the subject is wrong, regenerating with a different render style won't help — and the inverse is also true. Knowing which piece is broken is half the work.
Settings comparison
Here's a quick reference of the settings we used across the iterations above. You don't need to memorise these — you need to know they exist.
| Iteration | Model | CFG | Steps | Sampler |
|---|---|---|---|---|
| v1 — base | Photorealism | 7.0 | 30 | Euler a |
| v2 — softer | Photorealism | 5.5 | 30 | DPM++ 2M Karras |
| v3 — cinematic | Cinematic | 7.5 | 40 | DPM++ SDE Karras |
In the wild
We picked five recent community pieces and broke them down by layer. The common thread: the strongest pieces all spend more words on settingthan on render. The light is doing the heavy lifting.
- Specific lighting verbs beat generic adjectives.
- One concrete location beats three vague ones.
- Render style is the seasoning, not the meal.
Takeaways
You don't need to overhaul your prompt habits — just notice when an output disappoints you, and ask which of the four layers is at fault. Adjust that one. Regenerate. Repeat.
Next month we'll publish a follow-up on negative prompts: which kinds actually move the needle, and which are folklore.