Level 6Lesson 49⏱️ 45 min

AI Image Generation Fundamentals

Write prompts that actually produce the image you have in your head - structure, parameters, and the leading tools

The Anatomy of an Image Prompt

A weak image prompt is a single noun: "a dog". A strong one describes subject, style, composition, lighting, and detail in a deliberate order. Image models weight the front of the prompt more heavily, so lead with what matters most.

[SUBJECT] + [DESCRIPTORS] + [STYLE/MEDIUM] + [COMPOSITION] + [LIGHTING] + [QUALITY]

Example:
"A golden retriever puppy sitting in tall grass,
 photorealistic, shallow depth of field,
 shot on 85mm lens, soft golden-hour backlight,
 highly detailed, 4k"
Order matters. "Photorealistic puppy" and "puppy, photorealistic" can differ. Put the single most important word first.

The Five Building Blocks

1 - Subject

Who/what is in the frame. Be concrete: "a weathered fisherman" beats "a person".

2 - Style / medium

Photo, oil painting, 3D render, watercolor, line art, isometric, pixel art. This single choice changes everything.

3 - Composition

Close-up, wide shot, bird's-eye view, rule of thirds, centered, portrait vs landscape orientation.

4 - Lighting & mood

Golden hour, soft studio light, neon, dramatic shadows, overcast. Lighting is what makes images feel professional.

5 - Quality & technical cues

"Highly detailed", "8k", lens type, film stock. For photos, camera language ("shot on 50mm, f/1.8") nudges realism.

Parameters: Aspect Ratio, Seed, Negative Prompts

Beyond words, most tools expose controls. In Midjourney these are flags; in others they're UI fields. The concepts are universal.

--ar 16:9      Aspect ratio (16:9 video frame, 1:1 social, 9:16 reels)
--seed 12345   Fix the random seed → reproducible / iterable result
--no text      Negative prompt → exclude unwanted elements
--stylize 250  Stylization strength (tool-specific)
--chaos 30     Variety across the 4 results (tool-specific)
Workflow tip: Once you get a result you like, grab its seed and reuse it while tweaking one descriptor at a time. This turns random luck into directed editing.

The Leading Image Tools (2026) - and When to Use Each

GPT Image 2     Best all-rounder. Strong prompt fidelity, in-image text,
                conversational editing ("now make the sky orange").

Midjourney V7   The aesthetic leader. Gorgeous, stylized, opinionated look.
                Great for concept art, moodboards, striking visuals.

FLUX            Photorealism king. Skin, hands, fabric that pass as photos.
                Runs locally (open weights) or via API.

Adobe Firefly   Trained ONLY on licensed/owned content → commercially safe.
                The default when you must sell or publish without legal risk.

Ideogram 3      Best at TEXT inside images - logos, posters, ads, signage.
Need words in your image? (a poster, a mockup, a logo) - reach for Ideogram or GPT Image 2. Pure diffusion models still mangle long text.

Iteration Loop: From Rough to Right

Nobody nails it on prompt #1. The professional loop:

Step 1 - Generate a batch of 4

Read which one is closest to your intent, even if none is perfect.

Step 2 - Create variations of the best one

Most tools have a "vary" / "remix" button that explores around a chosen image.

Step 3 - Lock the seed, change one thing

Adjust a single descriptor (lighting, color, angle) so changes are controlled, not random.

Step 4 - Upscale the winner

Run a final upscale for resolution and detail (covered next lesson).

Hands-On: Build a Prompt in Layers

Hands-on (20 min): Start with just "a coffee cup" and generate. Then regenerate, adding one layer each time: medium ("product photo") → composition ("top-down flat lay") → lighting ("soft window light") → detail ("steam rising, shallow depth of field, highly detailed"). Watch the image transform at each step. Save all five so you can see exactly what each building block contributed. This is the single best exercise for internalizing prompt structure.
Lesson 49 Quick Reference
Prompt order

Subject → descriptors → style → composition → lighting → quality; front-load what matters most

Style/medium

The highest-leverage word - photo vs painting vs 3D render changes the entire image

Aspect ratio

16:9 video, 1:1 social, 9:16 vertical/reels - set before generating

Camera language

"Shot on 85mm, f/1.8, golden hour" pushes images toward photorealism

Vary / remix

Explore around a chosen image instead of starting over

Text in images

Use Ideogram 3 or GPT Image 2; pure diffusion still struggles with words

Commercial-safe choice

Adobe Firefly - licensed training data, lowest legal risk for published work

← The Generative Media Landscape
Unlocks in ~12 min of reading