AI Image Generation Guide

How It Works

Describe the subject, desired style or reference, and final use (social post, website hero, product mockup, print, etc.). The guide selects the right model (Midjourney, DALL-E 3, Flux, Stable Diffusion, etc.), gives a copy-paste prompt template with all the high-leverage controls, aspect ratios and settings per platform, negative prompt suggestions, and a systematic way to iterate from rough to final so you reach a usable image in 3–8 generations instead of 30+.

What to Provide

Image Generation Blueprint

This is the finished deliverable.

1. Model / Tool Recommendation

Primary recommendation: [[Midjourney v6 / Flux.1 / DALL-E 3 / SD3 or specific hosted version]]

Why for your use: [[Best prompt adherence + style control / speed / cost / upscaling quality / commercial rights]]

Runner-up / fallback: [[Alternative tool for the same prompt if primary is busy or limited]]

When to switch tools mid-project:

• Need transparent background or precise product placement → dedicated product tools or post-process in Photoshop.
• Need perfect text in image → DALL-E 3 or Ideogram currently strongest.
• Need heavy customization / fine control → local SD or ComfyUI with specific models.

2. Prompt Formula (Use This Structure Every Time)

Subject + details: the main thing, with specifics (age, clothing, pose, expression, material, lighting).

Style / aesthetic: photography style, art movement, artist references, medium (oil, 3D render, vector, film still).

Composition & camera: framing (close-up, wide), angle (low, eye level, Dutch), lens (35mm, 85mm, macro), depth of field.

Lighting & mood: time of day, light direction, color temperature, atmosphere (foggy, cinematic, bright minimal).

Technical / quality boosters: "highly detailed, sharp focus, 8k, intricate" etc. Use sparingly and only what the model responds to.

Parameters (platform specific):

--ar [[width:height]] (e.g. 16:9, 3:2, 1:1, 9:16)

--stylize [[value]] or equivalent

--v [[version]] or model choice

Negative prompt (what to avoid):

blurry, low quality, text, watermark, deformed hands, extra limbs, ugly, oversaturated, cartoonish (unless wanted), [[your specific dislikes]]

3. Platform Sizing and Export Cheat Sheet

Generate at 2x final size when possible, then downscale for sharpness.

4. Ready-to-Use Prompt Examples

Example A: Clean product hero (e-commerce)

"Minimalist product photograph of a matte black wireless earbud case on a light gray seamless background, soft even studio lighting from top-left, subtle reflection on surface, sharp focus on the case edges, clean composition with negative space on the right, modern commercial photography, highly detailed, 8k --ar 1:1 --stylize 150"

Negative: "text, logo, people, hands, shadows that obscure product, cluttered, colorful background"

Example B: Lifestyle / aspirational (social)

"Confident young woman in her late 20s with natural curly hair, wearing a cream oversized blazer and silver hoops, standing in a bright sunlit loft with large windows and plants, warm golden hour light, soft bokeh background, candid confident expression looking slightly off camera, shot on 85mm lens f/1.8, film photography aesthetic, rich but natural colors --ar 4:5"

Example C: Conceptual / illustration (marketing)

"Illustration of a small team of diverse professionals building a bridge made of glowing data connections over a river of abstract problems, flat design with bold geometric shapes, limited palette of deep blue, teal, and warm orange accents, clean vector style with subtle grain, balanced composition, modern tech branding illustration --ar 16:9"

5. Iteration Workflow (3–8 Generations Max)

Pass 1: Rough direction (2–4 gens)

Use a shorter prompt focused on subject + style + --ar.

Pick the composition and general mood you like. Note what is wrong (wrong pose, bad lighting, extra limbs).

Pass 2: Targeted fixes (use remix / vary / inpaint)

• Wrong pose or hands: Inpaint or use --cref + strong subject reference + describe pose explicitly.
• Lighting wrong: Add "soft diffused window light from left, gentle shadows" or specific time of day.
• Composition off: Add "rule of thirds, negative space on right, subject slightly left of center".

Pass 3: Polish and upscale

Add quality boosters only if the model responds well.

Upscale the best one (Midjourney upscale, or external upscaler like Topaz / Magnific for fine detail).

Pass 4: Post-processing (almost always)

• Minor cleanup in Photoshop (remove artifacts, adjust color, add text safely).
• Background extension or generative fill if needed.
• Color grade to match brand.

6. Prompt Engineering Tips That Actually Work

• Be specific about camera, lens, and lighting before style words.
• Use artist or photography references the model knows ("shot on Arri Alexa, Gregory Crewdson lighting, Alphonse Mucha").
• Put the most important thing early in the prompt.
• For character consistency across multiple images: use the same seed + character reference or describe signature features every time.
• For product shots: describe material and finish precisely ("matte black plastic with subtle texture, brushed aluminum accents").
• Negative prompt is powerful: use it.

7. Common Failure Modes and Fixes

• Deformed hands / faces: Inpaint the hands, add "anatomically correct hands, natural fingers" or use a model strong at figures.
• Text in image unwanted: Strong negative prompt + "no text, no letters".
• Colors off-brand: Add exact hex references or "brand color palette deep navy and warm gold".
• Wrong aspect for platform: Always set --ar at generation time.
• Repetitive or boring: Add "unexpected angle" or "candid moment" or specific emotion.

Worked Examples (End-to-End)

Example 1: E-commerce Product on Clean Background

Subject: Matte ceramic mug with subtle logo.

Prompt evolution shown from v1 (too busy) → v3 (clean, correct lighting, good reflection).

Final: Used with transparent version generated separately + slight post color match.

Example 2: Social Carousel Hero Image (Lifestyle)

3 consistent images of the same person in different outfits for a 3-slide carousel.

Used character reference + same lighting description across generations.

Example 3: Conceptual Marketing Illustration

Abstract data-to-insight visual for a B2B landing page.

Prompt used geometric style references + limited palette. Generated 6, picked 1, upscaled, minor vector cleanup in Figma.

Format Checklist

Quick Reference Card

1. Subject + details first.
2. Style + camera + lighting.
3. --ar and key params.
4. Negative prompt what you hate.
5. Iterate surgically (one variable at a time).
6. Always post-process for final use.

Generate with intention and you will reach usable images fast. Update the [[tokens]] for every new project and keep a personal library of prompts that worked for your style.