Generate AI Images — Visual Thinking Through Words

OUTCOME — What You'll Have When You're Done

Three things will be in your hands after this tutorial.

What gets built

• One AI image-tool account logged in
• Three actual images you generated using the 5-element prompt (subject · style · lighting · composition · aspect ratio)
• One iteration cycle where you refined a result using follow-up messages, inpainting, or a reference image

📖 Term: What is a Diffusion Model? Most 2026 AI image tools run on diffusion models. They start from pure noise and, using what they've learned from billions of images, gradually subtract noise until a picture emerges. That's why the same prompt twice produces slightly different results — the starting noise differs every time.

Tutorial meta

Item	Value
Time	30 min (fast: 20 min, unhurried: 45 min)
Level	Beginner
Tools	Web browser + one AI-tool account
Design knowledge	None required
Photoshop/Figma	Not needed
Payment	Optional (completable on free tier)

Why this matters — in visual-thinking terms

The sticking point for non-designers is: "the image is in my head but my hand can't draw." The 2026 answer is not the hand but the mouth. Once you can describe the image in your head precisely in words, that sentence IS the picture. This tutorial doesn't teach design. It trains you in "breaking visual thinking into sentences." Before this 30-minute training, you write "draw a cute cat" and blame the tool when the result looks off. After, you write "a black cat sitting on a Korean hanok wooden floor, backlit golden-hour sun, low-angle closeup, 16:9, warm analog film tone, soft shadows" and pull a usable result on the first try.

Where you are now: You've seen what this tutorial delivers. Next, a one-minute prerequisites check.

---

PREREQUISITES — What You Need

Three things.

Required

• Web browser: Chrome, Safari, Firefox, Edge — any recent version
• One AI account capable of image generation: a ChatGPT account (DALL-E 3 and GPT-5 Image 2 built in), a Google account (Gemini's Nano Banana Pro built in), a Midjourney account, or a Flux account
• One folder to save results: create an ai-images folder on your desktop; it'll save you time for the full 30 minutes

📖 Term: Prompt Engineering for Images "Prompt engineering" is the craft of refining the sentence you send the AI to get better results. A text AI prompt is a sentence of logic (context + instruction + format). An image AI prompt is a sentence of vision (subject + style + lighting + composition + aspect ratio). These have different grammars. People who are strong on text-AI prompts often get stuck on image prompts at first. Same word, different language.

Not needed

• Photoshop, Figma, or other pro design software
• Drawing ability, color theory, composition knowledge
• A tablet or stylus
• A design degree

Environment check (1 minute)

If you're logged into any one of the following, you can start right away:

• chatgpt.com (DALL-E 3 and GPT-5 Image 2 built in; partial free tier)
• gemini.google.com (Nano Banana Pro built in — fast and cheap)
• midjourney.com (as of 2026, you can generate on the web without Discord)
• flux.ai or Replicate (Flux 2 / Flux Pro — best for text rendering)

If you have zero AI accounts, finish the Pick Your First AI Tool tutorial in 20 minutes first, then come back.

Where you are now: Environment ready. Next, a 2-minute WHY on why image prompts are a different grammar from text prompts.

---

WHY — Why Image Prompts Are "Visual Thinking"

If you're fluent with text AI, your first image generation will likely surprise you. "Write me a marketing plan" worked. Why does "make me a pretty image" fail? Different grammar.

Reason 1 — In images, there is no "information that can be left out in words."

When writing text, words like "nicely" or "professionally" carry dozens of implied nuances. The reader fills them in. Images can't do that. To draw a "pretty cat," the AI must decide: coat color, background, time of day, lighting, camera angle, artistic style, aspect ratio. If you don't tell it, the AI picks carelessly for you. That careless pick is "the bland stock photo." The image in your head is visible only to you. The AI can't see it. You have to draw it fully in words and hand it over.

Reason 2 — The AI converges on "average."

Diffusion models are trained by averaging billions of images. So if you ask with no guidance, you get the most common average. "Cat" = orange tabby, seated frontal pose, cafe or couch background — the average of the most common cat photos on the internet. Unless you plant concrete elements that deviate from average in your prompt, your result blends into hundreds of thousands of similar stock images.

The image in your head is yours alone. The AI can't see it. Draw it fully in words.

Reason 3 — Most "bad results" are not tool problems. They are instruction shortages.

When students show me results and say "this tool is no good," 90% of the time the prompt is under 15 characters. "puppy illustration," "website background," "thumbnail image." Not one of the five elements — subject, style, lighting, composition, aspect ratio — is present. Of course the tool gave a bland result. Same tool, same 15 seconds; a 100-character prompt vs a 400-character prompt produces 10× different quality.

Where you are now: You understand why image prompts are "visual thinking translated into sentences." Now to the hands-on: pick a tool, write a 5-element prompt, pull three actual images.

---

STEPS — 5 Hands-On Steps

Step 1. Map of 2026 image tools (3 min)

As of 2026, the image generation landscape looks like this. Each tool has different strengths.

Tool	How to access	Strength	Best for
DALL-E 3	Built into ChatGPT	Best text comprehension, strong with Korean	Beginners, starting in Korean
GPT-5 Image 2	ChatGPT Plus/Pro	Flagship quality; editing, inpaint, reference image integrated	Highest-quality work
Midjourney 7+	midjourney.com (web, no Discord)	Top artistic quality; excellent reference image / identity lock	Posters, artwork, branding
Nano Banana Pro	Built into Gemini app	Fastest, cheapest	Bulk generation, blog/SNS workflows
Flux 2 / Flux Pro	flux.ai, Replicate	Best text (letters) rendering; open-source based	Logos, typography, custom
Stable Diffusion 3.5 / SDXL	Local install or web UI	Infinite custom models; fully free	Power users, pros

First-choice guide

• Want to write comfortably in Korean → DALL-E 3 (ChatGPT)
• Want artistic, poster-grade results → Midjourney 7+
• Generate many thumbnails/blog images in bulk → Nano Banana Pro (Gemini)
• Posters or logos with text → Flux 2 / Flux Pro
• Need the highest quality with editing and reference image workflows → GPT-5 Image 2
• Developer, want full customization → Stable Diffusion 3.5 / SDXL

If you can't decide, start with DALL-E 3. If you already have a ChatGPT account you can use it with zero extra sign-up.

Step 2. Internalize the 5-element prompt (5 min)

Every image prompt splits into five elements.

Element 1 — Subject: What is in the frame

• A person? animal? object? landscape?
• How many? where? doing what?
• e.g., "a black cat sitting on a hanok wooden floor," "a coffee cup and laptop on a desk"

Element 2 — Style: In what artistic manner

• Photo, illustration, watercolor, oil painting, 3D render, pixel art, flat design...
• You can cite specific artists/movements — "Ghibli style," "Banksy-esque"
• e.g., "watercolor illustration," "cinematic photo," "minimal flat design"

📖 Term: Reference Image — the 2026 standard When words alone can't capture the style you want, attach a reference image. DALL-E 3, GPT-5 Image 2, Midjourney 7+, Nano Banana Pro, and Flux 2 all support image input as of 2026. Just say "same feel as this reference, but with subject XX." Midjourney 7+ and GPT-5 Image 2 have strengthened identity lock — you can place the same character or person consistently across different scenes. One caveat: don't use copyrighted images as references — stick to your own, public-domain, or explicitly licensed images.

Element 3 — Lighting: Under what light

• Backlit, side-lit, front-lit, softbox, golden hour, blue hour, studio, neon...
• Lighting is 80% of mood
• e.g., "backlit golden hour," "soft natural light," "neon cyberpunk lighting"

Element 4 — Composition: At what angle

• Low-angle, high-angle, frontal, closeup, wide shot, over-the-shoulder...
• "What is in the frame, and how big"
• e.g., "low-angle closeup," "wide-angle wide shot," "frontal half-body"

Element 5 — Aspect Ratio: In what canvas

• 16:9 (YouTube thumbnails, blog heroes)
• 9:16 (Instagram Stories, Reels, Shorts)
• 1:1 (Instagram feed)
• 3:2 (photo prints)
• Each tool sets this differently — ChatGPT accepts "16:9" written into the sentence; Midjourney uses the --ar 16:9 flag

📖 Term: Aspect Ratio vs Resolution Aspect ratio is the shape of the canvas (width:height). 16:9 is a wide rectangle; 1:1 is a square. Size doesn't matter. Resolution is the number of pixels inside that canvas. In image prompts, lock the aspect ratio first. Aspect ratio decides layout; resolution is a separate export setting.

Weave all five:

A black cat sitting on a hanok wooden floor, (subject)
warm analog film photo style, (style)
backlit golden-hour sun, (lighting)
low-angle closeup, (composition)
16:9 aspect ratio (aspect)

Splitting with commas is a pro tip. The AI interprets each element independently.

Step 3. Bad prompt vs good prompt — the real difference (5 min)

Open your chosen tool and run both prompts below, in order.

Bad prompt

draw me a cute cat

Result: an orange tabby seated frontal, cafe-ish background — extremely common composition. Ten people using this prompt get nine near-identical images.

Good prompt

A black cat sitting on a hanok wooden floor gazing into the distance,
warm analog film photo style (Kodak Portra 400 feel),
late-afternoon backlit golden-hour sun,
low-angle closeup, shallow depth of field,
softly blurred background, 16:9 aspect ratio

Result: dramatic backlight, sunlight catching the black fur, blurred hanok silhouette. Ready for a blog hero image on the spot.

Same tool, same 30 seconds, completely different output. That's the power of the 5-element prompt.

Step 4. Iterative refinement — follow-ups, inpainting, reference image (7 min)

Even with a good prompt, rarely do you hit 100% satisfaction on try one. Three techniques the 2026 tools support.

Technique 1 — Refine with follow-up messages

Treat it like a text conversation — iterate with follow-ups.

Pass 1: [image received]
Pass 2: The composition is good, but could the cat sit a bit more to the left? Make one hanok pillar in the background sharper.
Pass 3: Great. Same mood, but change the cat to a white cat.

Just as you chain follow-up questions in text chat, chain follow-up edits in image chat. Three or four rounds and the result leaps beyond your first try.

Technique 2 — Inpaint / Outpaint

When you want to change only a specific area. Supported by DALL-E 3, GPT-5 Image 2, Midjourney 7+, and Flux 2.

• Inpaint: draw a mask over the image and say "change only this part" — swap the cat's color while keeping the background, or remove an unwanted element
• Outpaint: extend the canvas beyond the image edge — turn a 16:9 image into 9:16 by expanding top and bottom, or generate the scene outside a photo's frame

📖 Term: Inpaint and Outpaint Inpainting redraws only a selected region of the image — fix distorted hands, remove an unwanted object, or swap a color. Outpainting extends the image beyond its current border, letting the AI fill in the expanded space naturally. Both are standard features in major 2026 tools, and they remove most of the need for a separate photo editor.

Technique 3 — Lock style and identity with a reference image

Once you have a result you like, attach it as a reference in your next prompt.

[attach the cat image you just generated]
Same cat character as in the reference — now place it at a snowy winter windowsill.

Midjourney 7+'s identity lock and GPT-5 Image 2's reference image support make it standard practice in 2026 to reuse a character or style consistently across multiple images.

📖 Term: Negative Prompts — less needed in 2026 Negative prompts used to be essential for blocking common defects like distorted hands or watermarks. The latest 2026 models (GPT-5 Image 2, Midjourney 7+, Flux 2) have greatly reduced these issues on their own. That said, if you want to explicitly exclude something, "no [X]" or "without [X]" in the main prompt still works. Stable Diffusion and Flux have a dedicated negative prompt field.

Step 5. Three real-use prompts (10 min)

Now pull three images you can actually use. Adapt the templates below to your own situation.

Use 1 — YouTube thumbnail (16:9)

[your face or signature character],
surprised expression, bold-contrast poster style,
neon lighting, frontal closeup upper body,
16:9 aspect ratio, highly readable background,
no text (text will be added later in the thumbnail editor)

Use 2 — Blog hero image (16:9)

[scene symbolizing your article — e.g., "laptop open on a desk with a warm coffee cup"],
minimal photography style, soft natural light,
overhead top-down composition, 16:9 aspect ratio,
warm beige tones,
no people, no text

Use 3 — Instagram Story (9:16)

[product or symbol to promote],
flat illustration style, brand color (centered on #3B82F6),
centered composition with room for a title at top,
9:16 aspect ratio,
no cluttered background, no multiple products

Save all three to the ai-images folder.

Where you are now: You've practiced the 5-element prompt, follow-up refinement, inpainting, and reference images; three usable images are in your folder. Next, a checklist to verify the process went right.

---

VERIFICATION — Confirming It's Done Right

If all four check out, you've completed this tutorial.

Check 1 — Account and first image

• Are you logged into your chosen image tool?
• Did you generate a bad prompt version AND a good prompt version for comparison?
• Did you see the difference with your own eyes?

Check 2 — 5-element prompt internalized

• Have you written one prompt containing all five elements (subject, style, lighting, composition, aspect ratio)?
• Did you separate them with commas?

Check 3 — Iterative refinement experience

• Did you use at least one of: follow-up message, inpaint/outpaint, or reference image to refine the result?

Check 4 — Three use-case images

• Are the three use-case images (thumbnail / hero / Story) saved in the ai-images folder?
• Is at least one of them high enough quality to actually use today?

Three out of four is a pass. You don't need all four on day one — the feel develops over time.

Where you are now: The basics of the 5-element prompt are in your hand. Next, variations that extend the same principle to other situations.

---

VARIATIONS — Extensions

Variation 1 — Product mockups

No product photos yet? Use AI mockups.

A white ceramic mug on a beige linen background,
with a blank "logo area" label space on the front (to be composited later),
soft side-lit studio lighting,
high-angle 45-degree composition, 3:2 aspect ratio,
no beverage, no human hands, no other props

Overlay your actual logo afterward in Photoshop or Canva. Or use Flux 2 to render the text directly inside the prompt.

Variation 2 — A stand-in character (instead of your face)

Many people feel uneasy showing their face. Make a symbolic character and use it consistently.

Late-30s male illustration character,
minimal flat design style,
navy blazer over white shirt,
frontal upper-body, centered,
white or beige background, 1:1 aspect ratio

Once a version you like exists, attach that image as a reference in every subsequent prompt and chain "same character as in the reference, now in situation XX." Midjourney 7+'s identity lock feature makes visual consistency across the whole channel easy to maintain.

Variation 3 — 10 SNS cards in one batch

Turning a blog post summary into 10 card images? Keep the prompt structure fixed and vary only the content.

Card #1: [key sentence 1], beige background, space for black sans-serif font, flat illustration, 1:1
Card #2: [key sentence 2], beige background, space for black sans-serif font, flat illustration, 1:1
...

Fix the structure, vary only the content, and 10 cards finish in 10 minutes. Add a bulk-generation tool like Canva on top and it's even faster.

Where you are now: You've seen three real-use variations of the 5-element prompt. Next, a troubleshooting guide for when you get stuck.

---

TROUBLE + NEXT — When You Get Stuck, and What Comes Next

Common problems

Problem 1 — "The result looks like a stock photo"

• Cause: prompt under 15 characters, or 3+ elements missing
• Fix: fill the missing elements in subject · style · lighting · composition · aspect ratio. Lighting and composition are the most frequently skipped

Problem 2 — "Hands come out wrong (6 fingers, distorted feet)"

• Cause: diffusion models still have weakness at hands/feet even in 2026 (though greatly improved in GPT-5 Image 2 and Midjourney 7+)
• Fix: (a) add "no distorted hands" to the prompt, (b) use a wide shot instead of closeup, (c) specify pose — hands in pockets, hands behind back, (d) use inpaint to regenerate just the problem area

Problem 3 — "Text comes out garbled"

• Cause: DALL-E 3 and Midjourney are weak at rendering text
• Fix: (a) switch to Flux 2 — the strongest for text rendering, or (b) generate with blank space where text should go, then add text in Canva or Photoshop

Problem 4 — "Same prompt, different results each time"

• Cause: diffusion models start from different noise each run (see section intro)
• Fix: this is normal. When you get a result you love, save it immediately, and if the tool offers "regenerate with same seed," use it (Midjourney, Flux 2 support this)

Problem 5 — "I wrote in Korean and got unrelated results"

• Cause: some tools understand Korean prompts less than English
• Fix: translate to English. Paste the Korean prompt into ChatGPT and ask "translate this into an English image prompt: [your Korean prompt]"; it returns an English version you can feed Midjourney or Flux 2

Next steps

With this tutorial, the basics of image generation are in your hand. Three paths forward.

Path A — Deepen prompt writing

• Take the Write Prompts That Work tutorial to learn the 3-element text prompt formula
• Mastering both text and image prompts multiplies your content speed by 10

Path B — Connect to a design system

• Once you've generated 10 images, register colors and fonts in Canva's Brand Kit and drop images into templates for a consistent series
• Canva Brand Kit tutorial (coming soon)

Path C — Extend into video

• Once images feel easy, move to video generation (Runway, Pika, Kling)
• The 5-element image prompt applies almost directly to video prompts

One last principle

The perfect image lives inside your head. The AI can't read it. It paints only as much as you pull out in sentences. "Not having design skills" actually means "not having practiced pulling the image in your head out as sentences." Repeat the 5-element prompt exercise 30 times, and in two weeks you'll be making thumbnails faster than your designer friend. This is the era of shooting pictures with words. The mouth is faster than the hand.

한옥 마루에 앉은 검은 고양이, (주제)
따뜻한 아날로그 필름 사진 스타일, (스타일)
역광 골든아워 햇살, (조명)
로우앵글 클로즈업, (구도)
16:9 비율 (비율)

귀여운 고양이 그려줘

한옥 마루에 앉아 먼 곳을 응시하는 검은 고양이,
따뜻한 아날로그 필름 사진 스타일 (코닥 포트라 400 느낌),
늦은 오후 역광 골든아워 햇살,
로우앵글 클로즈업, 얕은 심도,
부드럽게 흐려진 배경, 16:9 비율

1차 결과: [이미지 받음]
2차 요청: 구도는 좋은데 고양이가 조금 더 왼쪽에 있으면 좋겠어요. 배경의 한옥 기둥 하나가 더 선명하게 보이게 해주세요.
3차 요청: 좋습니다. 이제 같은 분위기로 고양이만 흰색 고양이로 바꿔주세요.

[이전에 생성한 고양이 이미지 첨부]
이 이미지와 동일한 고양이 캐릭터로, 이번엔 눈 내리는 겨울 창가 장면으로 바꿔주세요.

[당신의 얼굴 또는 상징 캐릭터],
놀란 표정, 선명한 대비와 포스터 스타일,
네온 조명, 정면 클로즈업 반신,
16:9 비율, 가독성 높은 배경,
no text (글자는 썸네일 편집기에서 직접 넣을 예정)

[글의 주제를 상징하는 장면 — 예: "책상 위에 펼쳐진 노트북과 따뜻한 커피잔"],
미니멀 포토그래피 스타일, 부드러운 자연광,
오버헤드 탑다운 구도, 16:9 비율,
따뜻한 베이지 톤,
no people, no text

[홍보할 제품 또는 상징 이미지],
플랫 일러스트 스타일, 브랜드 컬러(#3B82F6 중심),
중앙 배치 구도, 상단에 제목 영역 확보,
9:16 비율,
no cluttered background, no multiple products

베이지 린넨 배경에 놓인 흰색 세라믹 머그컵,
컵 앞면에 "로고 자리" 라벨 공간 확보 (나중에 합성할 예정),
부드러운 측광 스튜디오 조명,
하이앵글 45도 구도, 3:2 비율,
no beverage, no human hands, no other props

30대 후반 남성 일러스트 캐릭터,
미니멀 플랫 디자인 스타일,
네이비 블레이저에 흰 셔츠,
정면 상반신, 중앙 배치,
흰색 또는 베이지 배경, 1:1 비율

카드 #1: [핵심 문장 1], 베이지 배경, 검정 산세리프 폰트 자리, 플랫 일러스트, 1:1
카드 #2: [핵심 문장 2], 베이지 배경, 검정 산세리프 폰트 자리, 플랫 일러스트, 1:1
...