Back to Blog
AI Images

AI YouTube Thumbnails: How to Make Them, Step by Step

Make an AI YouTube thumbnail that reads on a phone: the 16:9 specs, a prompt recipe, a text-on-image workflow, and how to A/B test it in YouTube Studio.

9 min read
By OndaLabs Team

To make an AI YouTube thumbnail, pick one idea, generate a 16:9 image with little or no text, add your title in an editor, check the result at phone size, and export it to YouTube's specifications. Then test it against a second thumbnail in YouTube Studio.

AI YouTube thumbnail specs in 2026

YouTube's custom thumbnail help page sets the formats. Design for these, and you will not need to resize later.

| Format | Recommended size | Aspect ratio | Minimum | File size | |---|---|---|---|---| | Video | 3840 x 2160 | 16:9 | 640 px wide | 2 MB on mobile, 50 MB on desktop | | Shorts | 2160 x 3840 | 9:16 | 640 px high | 50 MB on desktop | | Podcast playlist | Square | 1:1 | Not listed | 10 MB on mobile, 50 MB on desktop |

Older guides still recommend 1280 x 720 and a 2 MB limit. A safe rule is to design at 16:9, export at least 1280 x 720, and keep the file under 2 MB if you upload from a phone. Two more details: uploading your own thumbnail needs a verified account, and Shorts thumbnails can only be added in YouTube Studio on a computer.

YouTube also limits uploads. Each channel has a daily custom-thumbnail upload limit, and the help page says you will see a message to try again in 24 hours if you reach it.

Step 1: Decide the one idea the thumbnail shows

Before you open an image tool, write the single thing the viewer should understand in under a second. "Me holding a broken microphone" is an idea. "Tips for better audio" is a topic. Thumbnails work when they show one idea, and the title adds the detail.

Step 2: Write the prompt

Describe the scene in plain sentences, then say what the image is for. Google's Gemini docs recommend stating the shot type, subject, setting, lighting and camera angle. For thumbnails, add one more instruction: where the empty space should be, so the title has room.

Google's docs describe a "negative space" pattern: one subject in part of the frame against an empty background, which the docs call "excellent for creating backgrounds ... where text will be overlaid". Here are three example prompts written for this article, not quoted from a vendor:

  • "A surprised podcaster in a dark studio, lit from the left by a warm lamp, looking at a microphone on the right third of the frame. Leave the left half empty and dark for a title."
  • "A close-up of a hand placing a small green plant on a wooden desk, shallow depth of field, bright window light from the top right. Empty cream background on the left for text."
  • "A cartoon-style rocket on a launch pad at sunset, bold flat colours, the rocket in the lower right, a clear blue sky on the left."

Keep the prompt to a few sentences. The OpenAI image prompting guide suggests a consistent order: scene, subject, key details, then constraints. For more on structure, read our prompt guide.

Step 3: Pick the model and aspect ratio

Use 16:9 for regular videos and 9:16 for Shorts. Which model you use depends on the ratio you need. In the OndaLabs image studio, the Google models (Nano Banana Lite, Nano Banana and Nano Banana Pro) offer 16:9 and 9:16. The GPT Image models offer only 1:1, 3:2 and 2:3, so they cannot produce a 16:9 frame in the studio.

Nano Banana goes up to 4K, which suits YouTube's 3840 x 2160 recommendation. In the studio it costs 2,000 credits at 1K, 3,000 at 2K and 6,000 at 4K. Nano Banana Pro costs 7,500 credits at 1K or 2K and 13,000 at 4K. The studio shows the cost before you generate.

Step 4: Add text the reliable way

You have two choices for text. You can ask the model to draw it, or you can add it afterwards in an editor. Both vendors list text as a limit. Google's docs also market "advanced text rendering" for marketing assets, but its Nano Banana 2.1 model card says small text and long paragraphs can render weakly. OpenAI says text placement and clarity can still struggle.

Our suggestion is to generate the image with no text, or with only a word or two, and add the title in an editor. You then control the font, size and position. If you do let the model write text, put the exact words in quotes, and name the font, size, colour and placement. OpenAI's cookbook suggests that same kind of detail.

For the title itself, vidIQ, a vendor that sells creator tools, suggests roughly three to five words in large, bold sans-serif type. That is vendor opinion, not YouTube policy.

Step 5: Check it at phone size

Most viewers will see your thumbnail on a phone. Check each draft this way:

  • Shrink the image to about the size of a phone thumbnail and see whether the subject still reads.
  • Keep faces and key text away from the edges, because phone feeds can crop a 16:9 thumbnail.
  • Keep text out of the lower-right corner, where the video's duration stamp appears.
  • Confirm the title and the image say the same thing.

What the data says about faces, contrast and text

vidIQ's 2026 thumbnail study looked at 500 breakout long-form videos across 30 niches, published in the month ending 30 June 2026. Its article is dated 14 July 2026, and the figures below come from it.

| Element | Share of breakout videos | |---|---| | Human face | 69% (75% of the top 100, 80% of the top 50) | | High-contrast colour | 56% | | Face or high-contrast colour | 89% | | Exaggerated expression | About 1 in 20 | | Median overlay text | 5 words |

Read these as a description of what successful videos looked like, not as proof that a face or text causes more clicks. vidIQ does not define "breakout", and it sells creator tools. Do not treat these numbers as a guaranteed lift. The vidIQ thumbnail design guide has the full study.

Step 6: Test and compare in YouTube Studio

YouTube's Test and compare feature lets you try up to three thumbnails or titles on a video. Some rules to know:

  • The winner is chosen by watch time, not click-through rate.
  • Each test should finish within two weeks. If there is no clear winner, the first thumbnail uploaded is used.
  • It runs only on a computer in YouTube Studio, with advanced features enabled.
  • It is not available for Shorts, scheduled Lives or Premieres. Premieres become eligible after they end.
  • Videos made for kids, mature-audience videos, private videos and age-restricted (18+) videos cannot be tested.
  • Changing the title or thumbnail during a test stops the test.

Rules to know: misleading thumbnails and AI disclosure

YouTube's spam policy lists "maliciously misleading titles, thumbnails, descriptions, or imagery to trick users" under malicious clickbait. Violations can lead to suspended monetisation, content removal with a strike, or termination. Three strikes within 90 days may end a channel, according to the spam policy page. A thumbnail that promises something the video does not deliver is the common risk.

On AI disclosure, YouTube's altered or synthetic content page lists using generative AI to create or improve a thumbnail under production assistance, which does not need the AI label. The label applies to realistic content, such as making a real person appear to say or do something they did not. The page does not address a realistic, AI-altered thumbnail directly, and it says its lists are not exhaustive. If your thumbnail shows a real person in a scene that did not happen, read the page closely and consider disclosing it.

Thumbnails must also follow YouTube's Community Guidelines. Nudity, hate speech, violence and harmful content can get a thumbnail rejected, and repeat offences can remove custom-thumbnail privileges for 30 days or end the account.

Make thumbnail art at OndaLabs

The OndaLabs image studio can generate 16:9 thumbnail backgrounds with the Google models and edit them with reference images. Image generation is included from the Starter plan, and commercial use is included on Starter and above. Add the title in your own editor. If your video needs a voiceover, the same account has text to speech.

For more on the prompts, read how to write AI image prompts that work. For model choice, see the best AI image generators in 2026.

Frequently asked questions

How do I make a YouTube thumbnail with AI?

Pick one idea, generate a 16:9 image with little or no text, and leave space for the title. Add the title in an editor, check the result at phone size, export it at 1280 x 720 or larger, and test it in YouTube Studio on a computer.

What size should a YouTube thumbnail be in 2026?

YouTube recommends 3840 x 2160 for regular videos, in 16:9, with a minimum width of 640 px. Shorts use 2160 x 3840 in 9:16. Keep the file under 2 MB if you upload from a phone.

Can AI generate readable text on a thumbnail?

Sometimes, but neither OpenAI nor Google presents it as reliable. Keep the text to a word or two, put exact words in quotes when you ask for them, and add the main title in an editor so you control the font and size.

Do I have to disclose an AI-generated thumbnail on YouTube?

YouTube's help page says that using generative AI for a thumbnail is production assistance, which does not need the label. The label applies to realistic content, such as a real person shown saying or doing something they did not. Read the page and decide for your case.

How do I A/B test YouTube thumbnails?

Use Test and compare in YouTube Studio on a computer. You can test up to three thumbnails or titles. The winner is chosen by watch time, not click-through rate, and each test should finish within two weeks.