We do not recommend a prompting technique unless we have actually tried it. This page describes how we test, so you know what stands behind the advice on ArtPrompts Generator. The short version: we generate real images across multiple models before we write, and we keep the prompts so our results can be reproduced.
We generate before we recommend
When we claim a keyword, structure or setting does something, it is because we watched it do that thing. A typical test starts with a question — does this style word change the composition or only the palette; does raising this value sharpen detail or just add contrast — and we answer it by running prompts and comparing the output. Guesswork and second-hand tips get verified before they reach a guide.
We test across models
The same words behave differently in different systems. A phrase that steers Midjourney cleanly may do little in Stable Diffusion, and DALL·E interprets natural language in its own way. Where a guide is meant to be general, we check the technique across more than one of the models we cover — Midjourney, Stable Diffusion and DALL·E in particular, and Flux or Niji where relevant — and we tell you where behaviour diverges rather than pretending one recipe fits all.
We change one thing at a time
To understand what a variable actually does, we hold the rest of the prompt steady and change only the element under test — one keyword, one parameter value, one ordering. Comparing near-identical prompts side by side is the only reliable way to separate a real effect from randomness, and it is how Kai pins down the technical claims in our guides.
We account for the seed
Image models are stochastic: the same prompt can yield different pictures from run to run. So we never judge a technique on a single image. We generate several, and where a model supports it we fix the seed to isolate the effect of a change from the noise of ordinary variation. If a result only appears once and never again, we don’t treat it as a reliable technique.
We keep the prompts
Every recommendation is meant to be reproducible. We record the prompts, the model and version, and the key settings behind our examples, so that what you read can, in principle, be recreated. This also keeps us honest: if we can’t reproduce a result ourselves, it doesn’t get published.
What we can’t promise
Two honest caveats. First, models change — a technique verified on one version may drift after an update, and we note versions where it matters (see our editorial policy). Second, because generation is random, your results will vary even with an identical prompt; we aim to show you what reliably shifts the odds in your favour, not to guarantee a specific picture. When you are ready to put this into practice, our generator guide walks you through building a prompt step by step.