How to Make a YouTube Thumbnail With AI: Size, Examples and a Free ChatGPT Prompt

1. Check your channel can upload and test

In YouTube Studio, check which features your channel has. You want intermediate features on (phone verification) for custom thumbnails, and advanced features on for Test & compare. The steps are on YouTube's verification page and advanced features page (checked 15 September 2026).

Why: if advanced features are off, you can still upload a thumbnail, but you can't run the test in step 12. Better to know on day 1.

2. Decide what the thumbnail says that the title doesn't

Title and thumbnail are 2 halves of one pitch. Don't put the same sentence in both.

If the title is "How I book 40 consults a month with one reel", the thumbnail shouldn't say "40 consults". Show the calendar, or the reel on a phone. Ask: what does the picture add? If nothing, you have a caption. The same thinking applies to the first line of a video, covered in how to write hooks for reels.

Why: if you can't design a thumbnail before you shoot, the video doesn't have an idea yet.

3. Find 2 outliers from different channels

Search YouTube for the phrase someone would type to find your video. The full method is in finding YouTube outlier videos. Note the first 20 results and their views, take the page median, and divide each video by it. You want the video that beat the ones around it, not the biggest channel. A 40,000-subscriber channel pulling 2 million views is a signal. A 10-million one doing 2 million is a quiet Tuesday.

For the top 3, write what idea each thumbnail promises, not what it shows. Screenshot the best 2 from different channels. The research prompt below does the math and drafts headlines.

Why: 1 reference gives 4 outputs that look alike. 2 from different channels give you 2 directions.

Here are the first twenty YouTube results for "[YOUR SEARCH TERM]" with their view counts: [PASTE].
Work out the MEDIAN of the page. Tell me which videos are the biggest outliers against that median and give the ratio for each. Ignore channel size entirely.
For the top three, tell me in one line what IDEA each thumbnail is selling, not what it shows.
Then here is my script or outline: [PASTE].
Propose five headlines of three words or fewer for my thumbnail. My title will be "[YOUR TITLE]", so every headline must add something the title does not already say. Reject any that restate it.
For each headline, name the single object that should sit beside me in the frame, specifically enough that an image model could not get it wrong.

4. Pick one layout: 4 annotated examples

Illustrative examples written for this guide. Not real client thumbnails, and no result is implied.

Example 1: face + 3 words. An agency owner explaining AI video ads. Left third: the owner, cut out, calm expression. Right two thirds: "Ads Without Actors" in a condensed font, "Without" in the accent color. Beside the face: a phone showing a vertical ad. Why it works: the face gives the eye a landing point, 3 words read at phone size, and the phone tells you what kind of ads.

Example 2: before and after. A clinic owner's video on cleaning up lead follow-up. Left half: a cluttered inbox, desaturated. Right half: one calendar with booked slots, full color. Text across the split: "Inbox to Calendar". Why it works: the change is the story, so it needs no face. Watch out: only show a before and after that's true, and never a patient's body without the rules for your field.

Example 3: object close-up. A video on text-message booking. The whole frame is a hand holding a phone with one text bubble, "Booked for Tuesday", and nothing else. Headline: "Booked By Text". Why it works: one object, big, is the easiest thing to read at 320 pixels.

Example 4: contrast split. A comparison video. 2 panels, red and blue, one tool logo each, a large "or" in the middle, text "Which One?". Why it works: 2 colors and 2 objects tell you it's a choice before you read a word. Only use logos the way each company allows.

Why: naming the layout in the prompt stops the model from stacking every idea into one busy frame.

5. Get your face on file, or skip your face

5 photos or more: straight on and slightly turned, good light, plain background, same era. Cut-outs on white work best. With 1 or 2, the model makes a plausible person who isn't you. Generate one and look before you trust anyone's claims about likeness. Why likeness drifts is covered in consistent AI characters.

Ask if your face belongs there. For a clinic, agency or B2B channel, the subject is often the result, the screen or the object, which is layouts 2 to 4.

Why: a stranger's face with your name on the channel does more damage than no face.

6. Write 3 words and name 1 object

Not 5 words. If it needs a 4th, it's a title. Control Your Agents. No Photoshoot Needed. Ads Without Actors. Count the words on your references first. It'll be fewer than you planned. Be wary of click-through stats on this topic: the ones I found in thumbnail blogs didn't link to a primary source.

Then name the second thing in the frame exactly. Not "something related to ads" but "an ads dashboard with one row ringed in red". A vague noun gets a vague object. One object, and at most 3 elements including you. More on writing instructions a model follows: how to write better AI prompts.

Why: the words and the object are the only parts of the prompt the model can't guess for you.

7. Run the free ChatGPT prompt, then fix by conversation

Open ChatGPT, start a new chat, attach 1 reference and your face photos, and paste the prompt below with your layout, 3 words and object filled in. Ask for 2 versions. Repeat in a fresh chat with the other reference. That's 4 candidates. Image limits depend on your ChatGPT plan, and OpenAI changes them, so check what your plan allows.

Edit, don't regenerate. Use the fix prompt and say "keep the rest the same", the same difference-plus-keep-list idea as Nano Banana image editing. Re-rolling throws away what worked.

Composition: cut out onto a graphic looks designed. Your real room looks realistic, as long as it passes the checks in making AI images look real. Expression: ask for calm, because big expressions are where the likeness breaks.

Why: a new chat per reference stops the model mixing ideas from the last thumbnail into this one.

You are designing a YouTube thumbnail. I am attaching a reference thumbnail from a video that beat its topic, and [N] photos of me.
Redesign the CONCEPT of the reference in my own style. Keep what makes it clickable, which is the hook and the composition idea. Change the colors, the typography and the subject treatment so it is original and could never be called a copy of the reference.
Layout: [face + 3 words / before and after / object close-up / contrast split].
The person in the thumbnail is me, from the photos: keep my exact face, do not beautify or restyle it.
The ONLY text is "[YOUR THREE WORDS]", enormous, bold condensed sans, one or two lines, filling about half the frame width, one word in an accent color. No other words anywhere, no logos, no watermark.
The video title is "[YOUR TITLE]". The thumbnail text must ADD to that title, not repeat it. If my three words restate the title, say so before you generate anything.
Include this as a second focal point beside me: [NAME THE OBJECT EXACTLY].
Composition: [cut me out and place me on a built graphic / OR: put me in this room, using the attached photo of my actual space].
Ground: [light off white / deep charcoal].
My expression: [calm and confident / warm smile / mid explanation]. Keep it moderate so the likeness holds.
Render one single full frame image in a wide 16:9 landscape shape, never a grid or collage.
Give me two versions that differ in layout and color.

# the fix prompt, same chat
Here is the thumbnail you made. Change only these things and keep the rest the same:
- the headline is [N] words, cut it to three
- the face is [wrong expression], make it [what you want], and keep it moderate so it still looks like me
- there is nothing to read besides the headline, add [the object]
Do not re-generate the image from scratch. Edit this one.
I will check the legibility at 320 pixels myself; do not tell me whether it is readable.

8. The alternatives: Gemini (Nano Banana) and Ask Studio

Gemini. Google's image models go by Nano Banana: Nano Banana 2 and Nano Banana Pro. Google's image generation docs list 16:9 as a supported ratio and resolutions up to 4K, and say every generated image carries an invisible SynthID watermark (checked 15 September 2026). Use the shorter prompt below in the Gemini app, then edit in the same chat.

Ask Studio. YouTube's thumbnail help page describes Ask Studio suggesting thumbnails based on your video's transcript and themes, and taking follow-ups like making the text larger. If you see it in your Studio, try it first. It's the one option that has read your video.

Why: run the same idea through 2 tools and you get different styles to test, not 4 near-copies.

[Attach 1 reference thumbnail and your face photos, with the face photos last.]
Create a 16:9 YouTube thumbnail. The final images I attached are the real person: keep their face exactly, no restyling.
Use the reference only for its idea and composition. Change colors, type and background so it is original.
Layout: [face + 3 words / before and after / object close-up / contrast split].
Only text: "[YOUR THREE WORDS]", huge bold condensed sans, one accent word. Nothing else written anywhere.
Second focal point: [NAME THE OBJECT EXACTLY].
Expression: [calm and confident]. One single image, not a grid.

# follow-up edit, same chat
Keep everything the same except: [one change].

9. Decide who renders the text

Let the model render text while exploring. You can change wording, font or color by asking, and misspellings are a one-line fix prompt.

Add it yourself in a real font when shipping. You control it, the look repeats across the channel, and you can change the headline at 11pm without losing the picture. Any free editor like Canva or Google Slides works; the three fonts below are the ones my script uses.

For a company channel the real font usually wins. A locked font, accent color and frame are a brand asset. A different AI typeface every week isn't.

Why: the text is the part you'll change after a test, so keep it editable.

# real-font headline pass: 3 fonts ship with my thumbnail script
ls ~/.claude/skills/youtube-thumbnails/assets/fonts/
#   Anton-Regular.ttf  ArchivoBlack-Regular.ttf  BebasNeue-Regular.ttf

python3 ~/.claude/skills/youtube-thumbnails/scripts/generate.py --help
#   --prompt --ref --job --out --slug --fit {pad,cover,preserve}
#   --open --out-size --quality --format {jpeg,png}

10. Judge it at phone size yourself: the 320 pixel check

Shrink every candidate to 320 pixels wide. Can't read the words or tell what the object is? It failed, however good it looks full size. 320 pixels isn't a YouTube spec; it's a quick stand-in for a small thumbnail in a phone feed or the suggested videos column.

Do this yourself. Don't ask the model to shrink its own output and grade it. It'll say yes. The terminal version takes one command; the no-terminal version is below it.

Why: most people see your thumbnail small, on a phone, for a second.

# optional, terminal: shrink a candidate to 320 pixels wide
python3 -c "
from PIL import Image
im = Image.open('thumb.jpg').convert('RGB')
im.resize((320, round(320*im.height/im.width)), Image.LANCZOS).save('thumb-320.jpg','JPEG',quality=92)"

# 1280x720 in gives 320x180 out.
# open thumb-320.jpg at 100% and read it from arm's length.

# no terminal? put the image in a Google Slides or Canva page,
# drag it to about the width of your thumb on a phone screen, and look from arm's length.

11. Export at the right size and upload

Save the winner as a 16:9 JPG. 3840 x 2160 is YouTube's recommendation; 1920 x 1080 is a lighter option that still clears the minimum. Check the file size: under 2 MB if you upload from a phone, under 50 MB from a computer (YouTube Help, checked 15 September 2026).

Upload in YouTube Studio: Content, pick the video, and under Thumbnail choose Upload file, then Save. It can take time for the new thumbnail to appear.

Why: "youtube thumbnail size too big" shows up in YouTube's own search suggestions, and a failed phone upload at 11pm is avoidable.

# optional, terminal: export a YouTube ready JPG
python3 -c "
from PIL import Image
im = Image.open('winner.png').convert('RGB')
im = im.resize((1920, 1080), Image.LANCZOS)   # or (3840, 2160) for the recommended size
im.save('winner-youtube.jpg', 'JPEG', quality=90)
import os; print(os.path.getsize('winner-youtube.jpg')//1024, 'KB')"

# over 2,000 KB and uploading from a phone? lower quality to 80 or upload from a computer (50 MB limit).

12. Let YouTube pick: run Test & compare

On a computer, open the video in YouTube Studio and start Test & compare. Add up to 3 thumbnails, titles, or both. YouTube's Test & compare page (checked 15 September 2026) says a test can take a few days and should finish within 2 weeks, and ends as a winner, performed same, or inconclusive, where the first option stays.

It picks on watch time share, not click-through rate. A thumbnail that over-promises can win clicks and still lose, which is the right outcome.

What you can't test: Shorts, private videos, videos made for kids or for mature audiences, and scheduled lives and Premieres until they've ended. Regular videos and live archives work.

After it ends, give the result to ChatGPT or Claude with the critique prompt, and feed what you learn into the next video's step 6. Keep the losers in the folder.

Why: your audience decides, not a cheat sheet, and not the model that drew it.

Here is a screenshot of the thumbnails I tested, with the Test & compare result: [winner / performed same / inconclusive], and the watch time share for each.
Tell me what the winner has that the others do not. Be specific about composition, contrast, where the eye lands first, and how many things compete for attention.
Then tell me which single change to my prompt would make the next set more like the winner.
Do not compliment the ones that lost.