How to Write an AI Image Prompt (and Why Midjourney Wants Commas While DALL-E Wants Sentences)

The problem: you typed "astronaut on Mars" and got a stock photo

You open Midjourney, type "astronaut on Mars," and get a washed-out figure in an orange suit on a flat red background. You open DALL-E 3 with the same prompt and get a cartoon. You open Stable Diffusion and get a distorted body with six fingers. The same five words produce three different disasters because each platform reads prompts in a different grammar — Midjourney reads comma-separated tags weighted by position, DALL-E 3 reads natural-language sentences weighted by descriptiveness, and Stable Diffusion reads underscore-joined tokens weighted by a separate negative prompt. A prompt that works on one platform produces garbage on another. The honest move is a generator that outputs four platform-specific versions from the same subject, art style, mood, and lighting — not one prompt pasted into four tools.

Fastest path

Open the AI Image Prompt Generator, type the subject, pick the art style, click Generate Prompts.

Subject:  a lone astronaut standing on a red desert planet at dusk
Style:    cinematic
Mood:     mysterious
Lighting: golden hour

→ Midjourney:    "a lone astronaut standing on a red desert planet at dusk,
                  Cinematic, cinematic photography, movie still, anamorphic
                  lens, mysterious mood, golden hour lighting --ar 16:9 --v 6 --q 2"
→ DALL-E 3:      "A highly detailed cinematic artwork of a lone astronaut..."
→ Stable Diff:   "Positive: a_lone_astronaut..., Cinematic, mysterious, golden_hour_lighting...
                  Negative: blurry, low quality, bad anatomy, extra fingers..."
→ Firefly:       "A cinematic artwork of a lone astronaut... mysterious atmosphere,
                  golden hour lighting. cinematic photography, movie still..."

The tool took five inputs and generated 3 variants (detailed, minimal, creative) for each of 4 platforms — 12 prompts total, each in the grammar its platform expects. The rest of this guide is why Midjourney uses commas while DALL-E uses sentences, why Stable Diffusion needs negative prompts, why the art style tag does more work than the subject, and why 3 variants beat 1.

The substance: four grammars, nine styles, three variants

The four prompt grammars

Each AI image platform has its own prompt grammar. The tool builds four separate prompts from the same inputs because a single prompt pasted across platforms produces inconsistent results.

Midjourney reads comma-separated tags, weighted by position and frequency. The first tag carries the most weight; the last tag is a suggestion. Parameters go after --: --ar 16:9 sets aspect ratio, --v 6 sets the model version, --q 2 sets quality, --style raw reduces the default Midjourney aesthetic. The tool's buildMidjourneyPrompt places the subject first, art style second, style tags third, then mood, lighting, and details, then the parameters. The positional weighting is why a weak subject ("a dog") with a strong style tag ("cinematic photography, anamorphic lens, directed by Christopher Nolan") produces a better image than a strong subject with no style — the style tags are where the visual quality lives.

DALL-E 3 reads natural-language sentences. It doesn't weight by position — it parses the sentence semantically and renders what it understands. "A highly detailed cinematic artwork of a lone astronaut standing on a red desert planet at dusk, illuminated by golden hour. The overall mood is mysterious. Ultra high resolution with fine details." DALL-E ignores Midjourney-style parameters (--ar, --q) and Stable Diffusion-style token weighting ((word:1.2)). The tool's buildDallePrompt writes full sentences with descriptive clauses because that's what DALL-E's language model expects.

Stable Diffusion reads underscore-joined tokens and uses a separate negative prompt. The positive prompt is a list of tags separated by commas, with spaces replaced by underscores (a_lone_astronaut, cinematic_photography, mysterious, golden_hour_lighting, best_quality, masterpiece, 8k). The negative prompt is a list of what to exclude — and this is where Stable Diffusion diverges from the other three. The tool's negativePromptBase is a fixed string: blurry, low quality, low resolution, bad anatomy, bad hands, extra fingers, missing fingers, distorted, deformed, ugly, watermark, text, signature, jpeg artifacts, noise, grainy. This negative prompt is why Stable Diffusion images don't have the six-finger problem — the negative prompt explicitly excludes the common failure modes.

Adobe Firefly reads descriptive sentences, optimized for commercial safety. Firefly is trained on licensed content (Adobe Stock, public domain), so it won't generate copyrighted characters or trademarked logos. The tool's buildFireflyPrompt writes sentences similar to DALL-E but shorter and more descriptive — Firefly rewards brevity and clarity over long tag lists.

The nine art styles and their modifier stacks

The art style is the single most impactful input — more than the subject, more than the mood. The tool's styleModifiers map carries six style tags, a quality flag, and a reference artist for each of nine styles:

Style	Quality	Reference artist	Key tags
Photorealistic	--q 2	Annie Leibovitz	hyperrealistic, DSLR quality, 8K
Digital Art	--q 2	Artgerm	digital illustration, concept art, artstation trending
Oil Painting	--q 2	Rembrandt	oil on canvas, impasto technique, old master style
Watercolor	--q 1	John Singer Sargent	soft washes, paper texture, loose brushwork
Anime	--q 2	Makoto Shinkai	anime style, cel shading, studio ghibli inspired
Sketch	--q 1	Leonardo da Vinci	pencil sketch, cross-hatching, graphite
3D Render	--q 2	Unreal Engine 5	Cinema 4D, octane render, subsurface scattering
Pixel Art	--q 1	—	16-bit sprite, isometric, limited color palette
Cinematic	--q 2	Christopher Nolan	movie still, anamorphic lens, dramatic lighting

The reference artist tag is the most powerful modifier in the stack. "In the style of Rembrandt" tells the model to borrow Rembrandt's lighting, brushwork, and composition — and it works because the model has seen Rembrandt's paintings in training. The quality flag (--q 1 vs --q 2) is Midjourney-specific: --q 2 doubles the rendering time for higher detail; --q 1 is faster but softer. Watercolor, sketch, and pixel art use --q 1 because their aesthetics are intentionally loose — --q 2 on a pencil sketch produces over-rendered shading that kills the hand-drawn look.

The three variants: detailed, minimal, creative

The tool generates three variants per platform because different use cases want different prompt densities:

• Detailed: every tag the style stack offers — subject, art style, 4 style tags, mood, lighting, details, quality modifiers, aspect ratio, version. Use this when you want maximum control and maximum quality.
• Minimal: subject, art style, one style tag, lighting, aspect ratio, quality. Use this when you want the model to fill in the blanks — minimal prompts give the model more creative latitude and often produce more interesting compositions.
• Creative: subject, reference artist, surreal composition, 2 style tags, mood, lighting, "hyper-detailed, trending on artstation," --style raw. Use this when you want something unexpected — the creative variant injects surreal and dreamlike tags that push the model away from literal interpretation.

The three variants exist because a single prompt is a coin flip. The detailed variant is the safe bet; the minimal variant is the creative bet; the creative variant is the wild bet. Generate all three, pick the one that worked, iterate from there.

The five lighting and five mood options

Lighting changes the image more than any other modifier except art style. The tool's lightingOptions map carries four tags per lighting type:

• Natural: golden hour, soft daylight, overcast sky
• Dramatic: chiaroscuro, high contrast, rim lighting, backlit
• Studio: three-point lighting, softbox, beauty lighting
• Night: neon lights, moonlight, city lights bokeh, candlelight
• Cinematic: god rays, volumetric fog, lens flare

Mood is subtler — four enhancer words per mood type (peaceful: serene, tranquil, calm; dramatic: intense, powerful, emotional; etc.). Mood affects the color palette and composition more than the lighting. "Mysterious" produces shadowy, low-key images with cool tones; "joyful" produces bright, high-key images with warm tones. The mood is injected into the prompt as a tag (mysterious mood in Midjourney) or a sentence (The overall mood is mysterious in DALL-E).

Gotchas

• Midjourney parameters don't work on DALL-E. --ar 16:9 --v 6 --q 2 are Midjourney-only. DALL-E 3, Stable Diffusion, and Firefly ignore them. If you paste a Midjourney prompt into DALL-E, the parameters become noise in the sentence. Use the platform-specific prompt.
• Stable Diffusion is the only platform that needs negative prompts. The negative prompt is where you exclude the six fingers, the bad anatomy, the watermark. DALL-E 3 and Midjourney v6 handle these failure modes internally — they don't need you to say "no extra fingers." Stable Diffusion does.
• The reference artist tag is the most powerful modifier. "In the style of Rembrandt" does more than any quality tag. But it's also a copyright gray area — Midjourney and Stable Diffusion have been sued over artist mimicry. If you're generating commercial work, use the style tags without the artist name, or use Firefly (trained on licensed content).
• The art style does more work than the subject. "A dog, cinematic photography, anamorphic lens, directed by Christopher Nolan" produces a better image than "a golden retriever running on a beach at sunset in the late afternoon light with waves crashing and seagulls flying overhead." The style tags carry the visual quality; the subject just sets the scene.
• Minimal prompts give more creative results than detailed prompts. Counterintuitive but true — a 5-tag prompt lets the model fill in the composition, while a 30-tag prompt over-constrains it. If your detailed prompts all look the same, try the minimal variant.
• The --q 1 flag is correct for watercolor, sketch, and pixel art. --q 2 on a pencil sketch produces over-rendered shading that kills the hand-drawn aesthetic. The tool sets --q 1 for these styles automatically — don't override it.
• The --style raw flag reduces the Midjourney house look. Midjourney v6 has a default aesthetic (oversaturated, slightly cartoonish, "Midjourney-looking"). --style raw turns it off and produces a more neutral, photographic result. The tool's creative variant uses --style raw for this reason.
• Firefly is the only commercially safe option. Firefly is trained on Adobe Stock and public domain content, so its output is indemnified against copyright claims. Midjourney, DALL-E, and Stable Diffusion are trained on scraped web content and carry legal risk for commercial use. If you're generating images for a client, use Firefly.
• The 700ms loading delay is artificial. The tool computes the prompts synchronously — the spinner is there because users don't trust instant results. The prompts are generated client-side; no data is sent to any server.

Summary

• Four platforms, four prompt grammars. Midjourney reads comma-separated tags with positional weighting and -- parameters. DALL-E 3 reads natural-language sentences and ignores parameters. Stable Diffusion reads underscore-joined tokens with a separate negative prompt. Firefly reads descriptive sentences optimized for commercial safety. The tool generates all four from the same inputs.
• The negative prompt is Stable Diffusion's secret. blurry, low quality, bad anatomy, bad hands, extra fingers, missing fingers, distorted, deformed, ugly, watermark, text, signature — this is what prevents the six-finger problem. DALL-E and Midjourney handle these internally; Stable Diffusion needs you to say it.
• Nine art styles with reference artists. The art style does more work than the subject. "In the style of Rembrandt" tells the model to borrow Rembrandt's lighting and brushwork. The quality flag (--q 1 vs --q 2) is set per style — watercolor and sketch use --q 1 because their aesthetics are intentionally loose.
• Three variants: detailed, minimal, creative. Detailed is the safe bet (every tag). Minimal is the creative bet (5 tags, model fills in the blanks). Creative is the wild bet (surreal tags, --style raw). Generate all three, pick the winner, iterate.
• Lighting changes the image more than mood. Five lighting types (natural, dramatic, studio, night, cinematic) each carry four tags. "Golden hour" warms the image; "chiaroscuro" creates high-contrast shadow. Mood is subtler — it shifts the color palette and composition.
• Firefly is the only commercially safe option. Trained on licensed content, indemnified against copyright claims. Midjourney, DALL-E, and Stable Diffusion carry legal risk for commercial work.
• Generate at the AI Image Prompt Generator; for AI text generation use AI Story Generator, for rewriting use Article Rewriter, and for paraphrasing use Paraphraser.