How to Write Prompts for AI Image Generation

A prompt that works across most AI image generators names the subject specifically, then style, then lighting, then composition — in that order — since this four-part structure transfers reasonably well between Midjourney, DALL-E, Stable Diffusion, and other platforms, even though each has its own quirks.

The four-part structure that transfers across platforms

Subject, medium/style, lighting, composition. Naming the subject specifically first, then layering on a real style reference, lighting description, and composition detail, produces a genuinely predictable starting point on nearly any image generator, before you get into platform-specific parameters. This holds true whether you’re working in Midjourney, DALL-E, Stable Diffusion, or Gemini’s Nano Banana, even though the exact syntax and emphasis shifts between them.

Where platforms genuinely diverge

Word order weighting, parameter syntax (like Midjourney’s –ar and –style raw), and how literally each model interprets text vary meaningfully between platforms. A prompt that works well on one platform may need adjustment — not a full rewrite — when moved to another. Nano Banana specifically favors natural conversational sentences over the more tag-like phrasing that works fine on some other platforms, which is a genuine structural difference rather than just a stylistic preference.

Specific style references beat adjectives everywhere

“Watercolor illustration” or “35mm film photograph” gives any image model something concrete to act on. “Beautiful” or “stunning” describes a reaction, not a visual instruction, and underperforms across every platform tested.

This holds regardless of which specific model you’re using, which makes it one of the more reliably transferable pieces of image-prompting advice available — unlike parameter syntax or word-order weighting, which genuinely differ between tools, the vague-versus-specific distinction in style language matters the same way everywhere.

How lighting language transfers across platforms

Lighting description is one of the more universally understood elements across image generators — terms like “golden hour,” “soft diffused light,” “dramatic side lighting,” or “overcast daylight” tend to be interpreted similarly regardless of platform, since lighting vocabulary draws on shared photographic and cinematic convention that most models were trained on similarly. This makes lighting one of the safer elements to reuse verbatim when adapting a prompt from one platform to another, compared to style references or parameters, which need more platform-specific adjustment.

Composition language and its limits

Composition terms (“close-up,” “wide shot,” “rule of thirds,” “centered”) also transfer reasonably well, though the precision of interpretation varies — some models handle explicit compositional instruction more literally than others. If composition is critical to your result and a platform seems to be ignoring it, being more explicit and simpler (“the subject fills most of the frame” rather than a technical term like “extreme close-up”) sometimes produces more reliable results than assuming the technical vocabulary will be interpreted precisely.

Building a personal prompt template that adapts across tools

Given how much of this structure transfers, it’s worth building a personal template based on the four-part structure, then maintaining a short list of what needs adjusting per platform — Midjourney’s parameter flags, Nano Banana’s preference for flowing sentences, any platform-specific quirks you’ve noticed. This turns “learning a new image tool” into “adapting a template” rather than starting from scratch each time a new platform launches.

Common mistakes when adapting prompts across platforms

Copying a prompt verbatim between platforms without adjustment is the most common mistake — a prompt with heavy Midjourney-specific parameter syntax pasted directly into a different tool either gets ignored or misinterpreted, since that syntax isn’t universal. Stripping platform-specific syntax and keeping the underlying subject-style-lighting-composition content is usually the right first step when moving a prompt to a new tool.

Handling aspect ratio and output dimensions across platforms

Aspect ratio control differs meaningfully across platforms — some use a parameter flag directly in the prompt text, others use a separate settings control outside the prompt entirely. Knowing your target platform’s specific mechanism matters, since attempting to specify aspect ratio in prompt text on a platform that expects a separate setting simply won’t work, and the model has no way to interpret that instruction correctly.

Regardless of mechanism, deciding your intended aspect ratio before writing the rest of the prompt tends to produce better composition than treating it as an afterthought, since aspect ratio genuinely shapes what composition makes sense — a wide framing invites more environmental context than a tight square crop would.

Negative prompting availability and behavior differences

Not every platform supports explicit negative prompting (specifying what to exclude) the same way. Where available, negative prompts work best targeted at something you’ve actually observed the model adding unprompted, rather than a defensive list covering every possible unwanted element. Where negative prompting isn’t available as a separate mechanism, achieving a similar effect usually means being more explicit and specific in the positive prompt about what should be there, crowding out room for unwanted additions through specificity rather than explicit exclusion.

Working with reference images across different platforms

Many current image generators support using a reference image alongside a text prompt — for style transfer, character consistency, or composition guidance. The specific mechanism and terminology varies (style reference, image prompt, character reference), but the underlying principle transfers: a reference image communicates visual information that’s often difficult or impossible to fully capture in text alone, particularly for maintaining consistency across multiple generations of the same subject or style.

When switching between platforms for a project requiring visual consistency, checking whether your new platform supports an equivalent reference-image mechanism is worth doing before assuming text description alone will achieve the same consistency you got with reference-image support on a previous platform.

How model updates affect prompt reliability over time

Every major image generation platform updates its underlying model periodically, and prompts tested and confirmed working at one point can behave differently after an update — sometimes better, sometimes requiring adjustment. This is worth building into how you think about prompt reliability generally: a prompt that worked perfectly six months ago deserves re-verification before you rely on it for something important today, rather than assuming permanent stability. Re-testing periodically and dating your confirmed results is a genuinely useful habit, not just for your own reference but for anything you publish for others to use.

Choosing a platform based on your actual creative goal

Beyond prompting technique, the platforms themselves have genuinely different strengths worth matching to your goal rather than defaulting to whichever tool you already know. Some platforms lean toward painterly, stylized default output; others lean toward photorealism; some handle text rendering within images more reliably than others; some offer stronger reference-image and consistency features. If you’re regularly fighting a platform’s default tendencies to get a result outside its natural strengths, it’s worth asking whether a different platform’s defaults would align better with your actual goal, rather than continuing to fight the same friction indefinitely.

Applying iteration principles that hold across every platform

Regardless of which specific image generator you’re using, the iteration approach that works well elsewhere applies here too — when a result is close but not quite right, adjusting the specific element that’s off and regenerating tends to work better than a complete rewrite from scratch. This mirrors the same principle covered in Midjourney-specific prompting, and it holds regardless of platform since it reflects something general about how these models respond to incremental versus wholesale prompt changes. A useful diagnostic before rewriting entirely: can you point to one specific word or phrase likely responsible for the unwanted result? If so, adjust just that piece first before assuming the whole prompt needs reconsidering. This targeted approach saves time and preserves whatever was already working correctly in the original prompt, rather than risking a completely fresh combination that might lose those working elements entirely and require starting over.

Try it yourself

[subject, described specifically], [medium/style], [lighting description], [composition detail].

This prompt worked on [platform]: [paste it]. I’m switching to [different platform]. What elements should transfer directly versus need adjustment?

FAQ

For platform-specific technique, see the AI image prompt generator or browse platform-specific guides for Midjourney and DALL-E.