Prompt Engineering Guide for Perfect AI Visuals

Writing the right prompt can improve AI visual quality up to 10x. Professional prompt engineering techniques used by the Pam Istanbul team for global brands across automotive and luxury sectors on Midjourney and Stable Diffusion — including subject, style, lighting, composition and negative prompts.

• The right prompt structure can improve AI visual quality up to 10x
• Subject, style, lighting, composition, aspect ratio, negative prompt — a 6-layer systematic structure
• Includes real prompts from Pam Istanbul's automotive and luxury brand productions
• Prompt logic differs between Midjourney and Stable Diffusion

Prompt engineering is the single highest-leverage skill in AI visual production. In Pam Istanbul's studio, we've observed the same Midjourney account produce outputs that differ in quality by 500% depending on who is writing the prompt. The gap isn't tool access — it's structured thinking about what a model needs to produce great results. This guide shares the exact prompt architecture we use for global brands across automotive, fashion and luxury sectors, including annotated real-production examples.

The 5-Layer Prompt Architecture?

Every effective AI visual prompt contains five layers, each controlling a different dimension of the output. Layer 1, Subject: what is the primary element? Be specific. "A luxury perfume bottle" is worse than "a slender cylindrical amber glass perfume bottle with gold metallic cap." Layer 2, Style: what visual language? Reference real-world aesthetics: "editorial photography for Vogue", "Hasselblad medium format", "Bauhaus minimal design." Abstract terms like "beautiful" don't work — concrete aesthetic references do. Layer 3, Lighting: lighting defines mood more than any other element. "Soft diffused natural light from upper left" produces a fundamentally different result than "single hard spotlight from directly above." Be specific about direction, hardness, and color temperature. Layer 4, Composition: "extreme close-up macro", "three-quarter overhead angle", "wide establishing shot with subject in lower third." Composition decisions should match the platform: vertical crop for Instagram, horizontal for YouTube, square for e-commerce. Layer 5, Parameters: --ar (aspect ratio), --v (model version), --stylize (aesthetic strength 0-1000), --quality (render time 0.25-2), --seed (for reproducibility).

Real Production Prompts: Annotated Examples?

Cartier holiday campaign jewelry visual: "luxury diamond tennis bracelet on black matte velvet surface, single shaft of light from upper right creating sparkle caustics on the diamonds, extreme macro close-up, black background, editorial jewelry photography for Vogue Paris, ultra-sharp focus, dark and dramatic, --ar 4:5 --v 6.1 --style raw --q 2 --stylize 850". Analysis: "velvet surface" gives texture contrast to diamonds; "shaft of light creating sparkle caustics" is a specific lighting term that creates the crystal-light effect; "--style raw" removes Midjourney's default aesthetic processing for a more photographic result. Mercedes-Benz product visual: "silver Mercedes-Benz S-Class front three-quarter angle, luxury automotive photography, gradient studio background from charcoal to white, dramatic low-angle perspective, studio softbox lighting, ultra-realistic render quality, shot for automotive magazine --ar 16:9 --v 6.1 --q 2 --stylize 600". Note: "shot for automotive magazine" is the most efficient style reference for this category.

Negative Prompts: The Most Overlooked Quality Lever?

Negative prompts (--no in Midjourney, the negative prompt field in Stable Diffusion) tell the model what to avoid. This single addition improves output quality by 30-50% without any other change. Pam Istanbul's standard negative prompt set: "blurry, out of focus, low quality, jpeg artifacts, pixelated, watermark, text, logo, deformed, bad anatomy, extra limbs, mutated hands, ugly, duplicate, overexposed, oversaturated, flat lighting, artificial". Category-specific additions: luxury brand → "cheap, plastic, fake, tacky"; food photography → "unappetizing, gray, cold food look, artificial"; fashion → "bad fit, unflattering, wrinkled incorrectly". The standard negative prompt list should be included in every brand prompt template.

Advanced Techniques: Style Reference, Character Reference, and Seed Control?

Three advanced techniques that professional AI operators use in brand production. Style Reference (--sref): upload a reference image and add --sref [image URL] to transfer its visual style to your output. Essential for brand consistency — the reference "teaches" Midjourney your brand aesthetic without fine-tuning. Character Reference (--cref): maintains a consistent character appearance across multiple images. Critical for fashion editorial where the same model appears in multiple scenes. Seed Control (--seed [number]): the seed value determines the random noise starting point. Using the same seed with slight prompt variations produces controlled, comparable outputs — essential for testing variations of the same composition. In production, record the seed of any strong output and use it as the foundation for the visual series.

Building a Brand Prompt Library: The System Behind Consistency?

When multiple operators produce visuals for the same brand without a shared prompt library, the results diverge. Each person uses different words, different lighting descriptions, different style references — and the outputs look like they come from different brands. Pam Istanbul delivers every brand engagement with a version-controlled prompt library containing 50-200 templates organized by category (product, editorial, social media, campaign) and subcategory (hero shot, lifestyle, detail, video). Each template includes: the base prompt, standard negative prompt set, parameters, notes on when to use it, and example outputs. This library becomes the brand's most important AI production asset — it encodes the brand's visual language in machine-readable form.

Platform-Specific Parameters and Format Guide?

Each platform has optimal parameters. Instagram Feed (4:5): --ar 4:5, stylize 600-800 for editorial, 400-600 for product. Instagram Stories/Reels (9:16): --ar 9:16, keep subject centered or in upper third. LinkedIn (1:1 or 1.91:1): --ar 1:1, more corporate tone, --stylize 400-600. YouTube thumbnail (16:9): --ar 16:9, high contrast, bold composition. Print/OOH: start with --q 2 for maximum quality, plan for upscaling to print resolution. Amazon product: --ar 1:1, pure white background prompt, --q 2.

Prompt

A text-based instruction given to an AI image generation model

Negative prompt

A parameter defining what should be excluded from the visual output

Aspect ratio

The width-to-height ratio of an image (e.g. 16:9, 1:1, 4:5)

Style reference

A reference image given to AI that calibrates the aesthetic direction of the output

Weight

A value determining how dominant a particular term is within a prompt

Learning prompt engineering takes time — and every brand has a different learning curve. Pam Istanbul prepares your brand-specific prompt library and manages your entire production process from start to finish.

Blog · Get a quote