How to Make a YouTube Thumbnail with AI
A thumbnail has one job: win the click in the half-second before someone scrolls past. Here is the right size, what actually drives clicks, and how to generate one by describing it — free, no Photoshop.
July 25, 2026 · 6 min read
Contents
Quick answer
You make a YouTube thumbnail with AI by uploading a frame or a photo, then describing the thumbnail you want — “bold 3-word title top-left, my face on the right looking shocked, bright yellow background, high contrast.” An AI image tool renders it at 16:9, and you export at 1280×720. On Shotari the built-in model does this for free, and for thumbnails with readable on-image text, GPT Image 2 (5 credits) is the one to reach for.
The half-second rule
Your thumbnail competes at the size of a postage stamp on a phone. If the title isn’t readable and the subject isn’t clear at that scale, nothing else matters.
The specs that matter
Get these right first — a beautiful thumbnail at the wrong size gets cropped or rejected. If your source frame isn’t 16:9, expand it to fit with AI outpainting instead of cropping your subject out — how uncropping works covers what it costs and where it breaks.
| Spec | Value |
|---|---|
| Resolution | 1280 × 720 px (minimum 640 wide) |
| Aspect ratio | 16:9 — the same ratio every player and shelf uses |
| File size | Under 2 MB |
| Format | JPG, PNG, or WebP |
| Bottom-right corner | Keep it clear — the video duration stamp sits there |
On Shotari, 16:9 is the Landscape aspect ratio. It carries a small ×1.1 credit multiplier over square output, so a 5-credit generation is 6 credits at 16:9 — worth knowing before you run a batch of variations.
What actually earns the click
Thumbnails don’t get clicked because they’re pretty. They get clicked because they promise something in a glance. Three levers do almost all the work.
One readable idea
Three or four words maximum, in a heavy font, filling a big part of the frame. If it takes more than a glance to read, it’s already lost. The title on the thumbnail should not repeat the video title — it should add curiosity.
A face with an emotion
Human faces pull the eye, and an exaggerated expression — shock, delight, disbelief — outperforms a neutral one. Even a hand pointing at the subject helps direct attention.
Contrast that survives shrinking
Bright, saturated backgrounds and a clear subject edge keep the thumbnail legible at phone size. Busy, low-contrast images turn to mud in the feed.
How to make one
1. Start from a frame or a photo
Upload a screenshot from the video or a clean photo of yourself. Giving the AI a reference keeps your face and the subject accurate instead of inventing a stranger.
2. Describe the layout, not just the scene
Say where things go: “title text top-left, subject on the right third, arrow pointing at the product.” Layout direction is what separates a thumbnail from a generic image.
3. Generate at 16:9, then export 1280×720
Pick the Landscape ratio, generate a few variations, and download the strongest. Draft at standard resolution; only upscale the winner.
Getting readable title text
Readable words inside the image are where most AI thumbnails fall apart — you get garbled, half-formed letters. This is a model choice, not a prompt trick. GPT Image 2 renders short headlines cleanly, which is exactly what a thumbnail needs, so it’s the model to pick when the title has to live inside the image. Keep the text to three or four words; even the best model gets shakier with a full sentence.
If you want total control over the wording, generate the thumbnail background and subject with AI, leave space for the title, and add the text in any free editor afterward. Cleaner than fighting the model for a specific font.
Prompts to copy
Reaction / commentary video
YouTube thumbnail, 16:9, the person from the reference photo on the right third with a shocked expression, bold 3-word title 'YOU WON'T BELIEVE' top-left in heavy white font with black outline, bright red gradient background, high contrast, clean edges
Tutorial / how-to
YouTube thumbnail, 16:9, a clean product or screen on the left, a large arrow pointing at it, short title 'FULL GUIDE' top-right in bold yellow, soft studio background, crisp and legible at small size
Vlog / lifestyle
YouTube thumbnail, 16:9, the subject from the reference photo centered and smiling, warm cinematic lighting, a short 2-word caption bottom-left, shallow depth of field, vivid saturated color
Product review
YouTube thumbnail, 16:9, the product from the reference image large and centered on a bold split-color background, a small 'WORTH IT?' tag top-right, dramatic lighting, sharp product edges
Mistakes that kill CTR
- ⚠
Cramming in a full sentence
Unreadable at feed size and hard for any model to render. Three or four words, always.
- ⚠
Repeating the video title
The title text is already next to the thumbnail. Duplicating it wastes the frame — use the thumbnail to add curiosity, not restate.
- ⚠
A busy, low-contrast background
It looks fine full-size and turns to mud on a phone. Test every thumbnail by shrinking it to a thumbnail.
- ⚠
Clutter in the bottom-right
The duration stamp covers it. Keep anything important out of that corner.
- ⚠
Inventing a face instead of using yours
Generating without a reference gives you a stranger. Upload your photo so the channel stays recognizably you.
FAQ
What size should a YouTube thumbnail be?
1280 × 720 px, 16:9, under 2 MB, as JPG, PNG or WebP. Keep the bottom-right corner clear — the duration stamp sits there and will cover anything you put under it.
Can I make a YouTube thumbnail with AI for free?
Generating is free on Shotari Basic at 0 credits, and guests get two runs a day without an account. Downloading needs an account, and the free download is watermarked — which is disqualifying for a thumbnail specifically, so the file you upload to YouTube has to come from paid credits.
Why does 16:9 cost more credits than square?
Landscape and portrait ratios carry a ×1.1 multiplier that is rounded up, so a 5-credit model is 6 credits at 16:9. Since thumbnails are always 16:9, budget for the surcharge on every paid run — and do your composition tests on the free model, where it costs nothing either way.
Which model handles the title text best?
GPT Image 2 or GPT Image 1.5. Legible on-image text is the thing they are best at, and a thumbnail lives or dies on whether three words are readable at postage-stamp size. Expect to re-run: long strings still come back garbled more often than short ones.
My source frame is not 16:9 — do I crop it?
Not if the crop cuts your subject out. Extend it instead: uncropping keeps the original frame intact and paints the extra width, so you reach 16:9 without losing the face that makes the thumbnail work.
How many variations should I generate?
Three or four is usually enough to find the layout, and all of them should be free runs. Switch to a paid model once — for the version you are actually uploading. Running twenty paid variations is how people spend a few hundred credits answering a question the free model answers.
Make your thumbnail now
Upload a frame, describe the layout, and generate a few 16:9 options — free to start, no design tool.
Open the thumbnail maker