AI

YouTube Thumbnail Prompts for ChatGPT and Gemini: 6 We Tested and Scored

We ran six copy-and-paste thumbnail prompts on three real videos, in ChatGPT and in Gemini's Nano Banana, and scored all 22 pictures. The one-line prompt most people type made up a baker and a hiker who aren't in the videos. Remaking your…

Alex15 min read
Generate a thumbnail from your video

Disclosure: we make ThumbnailPeak, which has its own thumbnail generator, and we scored everything with ThumbnailPeak's own critique. Every picture and score below is from 1 November 2026. The three videos are their creators' own, published under a Creative Commons licence.

A thumbnail prompt is only as good as what you give the AI to work from. We tested six prompts, from the one line most people type to a full brief with your own photos attached, on three real YouTube videos. Here's what we found:

  • The one-liner invents. "Make a YouTube thumbnail for my video" made up a smiling baker who isn't the bread video's host, and a different hiker with a different dog for a hiking video.
  • Remaking your current thumbnail scored best: 90 in ChatGPT and 85 in Gemini, from a creator thumbnail that scored 85.
  • Your face, reacting, scored 87 in ChatGPT for a Minecraft video whose own thumbnail scored 78.
  • For "X vs Y" titles, a split prompt scored 84 and 85, above the creator's own.
  • A brief can backfire. Our own brief for the hiking video asked for the person and dog to look small against the storm. It scored 69 and 75, below the one-liner.
  • ChatGPT came back 3:2 every time, even when the prompt said 16:9. Nano Banana came back wide every time.

Every prompt is below, ready to copy. Replace the words in [brackets].

How we tested

We picked three videos published under the Creative Commons Attribution licence, each a different kind of thumbnail problem:

We ran each prompt once in each of two places and kept the first picture:

  • ChatGPT, through OpenAI's API in the way that comes closest to the app: a chat model with OpenAI's image tool, set to chatgpt-image-latest.
  • Gemini's Nano Banana 2, which Google's Gemini page lists as free, through Google's developer platform.

The apps add instructions of their own, so your results there may differ. Where a prompt attaches frames, they're YouTube's own stills from the video. Then we scored every picture in ThumbnailPeak's thumbnail analyzer with the video's title, and scored the creators' real thumbnails the same way: 78 for Minecraft, 85 for the ciabatta and 80 for the bothy.

The rules every good prompt below follows

These six lines did more than any clever wording. They come from what our critique kept flagging in our ChatGPT test and our Nano Banana test:

  1. Ask for 16:9 (1280×720). It works in Gemini. In ChatGPT it doesn't, so crop afterwards.
  2. Attach real pictures of what's in your video: frames, photos, your current thumbnail.
  3. Describe the subject as it really looks. Colour, size, what it is.
  4. At most 3 words of text, none from the title. YouTube prints your title under the thumbnail.
  5. No people, logos or objects that aren't in your pictures.
  6. Keep the bottom-right corner empty. YouTube puts the video's length there.

1. The one-liner (where most people start)

Make a YouTube thumbnail for my video "[your title]".

Scores: 72 to 84. It's quick, and ChatGPT's Minecraft try scored 84, above the creator's own. But all six pictures were made up from the title alone:

Three rows of thumbnails, one row per video. Minecraft: the creator's own split "$1 / $1,000,000" over a dirt hut and a stone fort, 78; ChatGPT's split with a scared character in a dirt hole and a blond character at a bunker door, 84; Nano Banana's with the whole title across the top over two shelters with characters, 77. Ciabatta: the creator's own, the host holding torn bread under "WHY SO EXPENSIVE?", 85; ChatGPT's loaf on a board under "BETTER THAN STORE-BOUGHT!", 77; Nano Banana's smiling woman holding two loaves under "HOMEMADE CIABATTA! BETTER THAN THE STORE!", 78. Mountain bothy: the creator's own selfie in a red jacket pointing at a stone hut, with a black dog and "THIS WAS GRIM!", 80; ChatGPT's soaked stranger hugging a brown-and-white dog in the rain, 83; Nano Banana's woman sitting in a hut at night with a phone and a brown dog, under "OVERNIGHT ALONE IN TINY BOTHY! THINGS WENT WRONG!", 72.

  • It invented people. Nano Banana drew a smiling baker who isn't Skip The Junk's host. The bothy video is one hiker and a black dog; ChatGPT drew a different hiker with a brown-and-white dog, and Nano Banana a different person again.
  • It repeated the title. All six put the title's words on the picture. Nano Banana's Minecraft try has the entire title across the top.

Use it to see ideas, not to make your thumbnail.

2. The full brief

Make a 16:9 YouTube thumbnail (1280×720) for my video "[your title]". The attached frames are from the video: [describe the subject exactly as it looks, and say if the person in them is you, so the face stays the same]. Show [the main thing, big and clear], [where in the frame]. [One line on the idea: a reveal, a question, a result.] Add at most 3 short words of bold text, and don't use any words from the title. Don't add people, logos or objects that aren't in the frames. Keep the bottom-right corner empty, because YouTube puts the video's length there.

Scores: 73 to 89. For the ciabatta, with three frames of the host in the kitchen attached and the bread described as "torn open so you can see the big airy holes inside", ChatGPT made "LOOK INSIDE" with the host holding the loaf: 89, one of the best pictures in the test. Nano Banana's "EASY. AIRY. AMAZING." scored 81.

For Minecraft it went wrong, and the frames were why. YouTube's stills from that video happen to show two nice houses, not the $1 dirt hut. "Show the builds exactly as they look" gave us two nice houses, and the "$1 vs $1,000,000" idea was gone. ChatGPT's scored 73. Nano Banana labelled them "CHEAP" and "EXPENSIVE" anyway, and copied a red block from the player's hand in the frame into its picture: 82.

Four thumbnails. Ciabatta: ChatGPT's host smiling and holding a torn-open loaf under "LOOK INSIDE", 89; Nano Banana's host holding the open loaf under "EASY. AIRY. AMAZING.", 81. Minecraft: ChatGPT's two houses in front of a tornado under "SURVIVE?", 73; Nano Banana's two houses split by a red line, labelled "CHEAP" and "EXPENSIVE", with a red block from the game's hand in the corner, 82.

Before you use it: attach the frames that show what your title promises, not just any frames. And look for things the AI copied from your screen: game menus, a hand, a cursor.

3. Before and after

For builds, makeovers, restorations and anything that ends looking different from how it started:

Make a 16:9 YouTube thumbnail (1280×720) for my video "[your title]". The attached frames are from the video: draw [the subject] exactly as it looks in them, [describe it]. Show the [before] and the finished [after] side by side, big and clear. Add at most 3 short words of bold text, and don't use any words from the title. Don't add people who aren't in the frames, and no logos. Keep the bottom-right corner empty, because YouTube puts the video's length there.

Scores: 83 to 91. We tested this one in our Nano Banana post on a go-kart rebuild. Nano Banana 2 went from 61 with the one-liner to 85, and Nano Banana 2 Lite from 79 to 91. ChatGPT's scored 83. The pictures are in that post.

4. The versus split

For any title with "vs", "$1 vs $1,000,000", "cheap vs expensive", "beginner vs pro":

Make a 16:9 YouTube thumbnail (1280×720) for my video "[your title]". Split it into two halves with a sharp diagonal line. Left half: [the first thing], with "[short label]" in big white letters above it. Right half: [the second thing], with "[short label]" in big gold letters above it. [What's behind both.] Make the right half brighter and more dramatic than the left. [The style, e.g. "Everything in Minecraft's blocky style, like the attached frames."] No other text, no people, no logos. Keep the bottom-right corner empty, because YouTube puts the video's length there.

Scores: 84 and 85, against 78 for the creator's own thumbnail, which uses the same idea.

Three thumbnails of the same idea. The creator's own: "$1" over a small dirt hut and "$1,000,000" over a stone fort, with a tornado and a red arrow, 78. ChatGPT's: "$1" over a dirt hut and "$1,000,000" over a lit mansion, split by lightning, with black bars at the sides, 84. Nano Banana's: "$1" over a dirt hut and "$1,000,000" over a stone mansion under a purple tornado, split by a white lightning bolt, 85.

Both made the cheap side small and plain and the expensive side big and lit, which is the whole point. The critique's one complaint about the creator's own was that "both builds occupy a small area": the AI versions made them bigger. The labels repeat the title, and that's the one place it's worth it: the numbers are the idea.

5. Your face, reacting

For channels where your face is part of the thumbnail. Attach a clear, well-lit photo of yourself, facing the camera:

Make a 16:9 YouTube thumbnail (1280×720) for my video "[your title]". The attached photo is me. Put me large on the left, from the chest up, cut out with a thin white outline. Keep my face exactly as it is and change only my expression to [shocked / laughing / worried]. I'm looking at the right side, where [the moment from your video, like the attached frames]. Add at most 3 short words of bold text, and don't use any words from the title. No logos. Keep the bottom-right corner empty, because YouTube puts the video's length there.

We didn't use a real creator's face for this. The photo is a stand-in we made with AI for our tests, not a real person.

Scores: 87 in ChatGPT, 78 in Nano Banana.

The stand-in photo of a young man in a green sweater, smiling at the camera. ChatGPT's thumbnail: the same man with eyebrows raised and mouth open, looking at a Minecraft mansion with a tornado behind it, under "WILL IT HOLD?", 87. Nano Banana's: the same man with a shocked open mouth next to a blocky tornado hitting a mansion, under "INSANE TORNADO STORM", 78.

Both kept the face recognisable and changed only the expression. ChatGPT's "WILL IT HOLD?" asks the question the video answers. Nano Banana's "INSANE TORNADO STORM" is three words of description, and the critique marked its title fit down to 64 because neither picture shows the $1-versus-$1,000,000 idea. A face prompt still needs the video's idea in it.

If you want your face in every thumbnail without writing this each time, our post on AI thumbnails with your face shows how a face library does it.

6. Remake my current thumbnail

The easiest prompt with the best result. Attach the thumbnail you already have:

This is my current thumbnail for my video "[your title]". Make a new version that's easier to read on a phone. Keep my face and [the main subject] exactly as they are. Make [the main subject] bigger and the background simpler and darker than it. Use at most 3 short words of bold text. Don't add anything that isn't already in my thumbnail. Make it 16:9 (1280×720), and keep the bottom-right corner empty, because YouTube puts the video's length there.

Scores: 90 in ChatGPT, 85 in Nano Banana, from a thumbnail that scored 85.

Three thumbnails. The creator's own: the host smiling beside a torn loaf under "WHY SO EXPENSIVE?", 85. ChatGPT's remake: the same host and loaf, bigger, on a dark background, under "WHY SO", 90. Nano Banana's remake: the host and the loaf on a dark background under "BEST HOMEMADE", 85.

It works because the AI has everything it needs: your face, your subject, your idea. It only has to make them clearer.

But read the words. Told to use at most 3 words, ChatGPT cut "WHY SO EXPENSIVE?" down to "WHY SO", which no longer means anything. Our critique still scored it 90, though its first note was that "The headline stops before naming what is unusual about the bread." A score can't stand in for reading the result yourself. Nano Banana swapped in "BEST HOMEMADE", which repeats the title.

A brief that went wrong, and the fix

For the bothy video we attached the creator's own thumbnail (the creator, the dog and the hut) and asked for "the tiny stone bothy on a wild, rainy hillside, with me and my dog in front of it, small next to the weather."

That last part was a mistake. Both tools did exactly what we asked: a storm, a hut, and a person and dog so small you can't make them out on a phone. ChatGPT's scored 69 and Nano Banana's 75. The critique said it plainly: "The person and dog are too small to make out clearly at 168×94."

So we asked for the opposite: "Put me and my dog big in the front, from the waist up, close to the camera". That brought them up to 77 and 78.

Four thumbnails. Asked for them small: ChatGPT's tiny hiker and dog walking to a hut in a storm under "SOAKED THROUGH", 69; Nano Banana's tiny hiker and dog beside a stone hut under "TOTAL NIGHTMARE", 75. Asked for them big: ChatGPT's hiker in a red hood beside a black dog, with the hut behind them under "NIGHTMARE WEATHER", 77; Nano Banana's hiker pointing at the hut with a black dog beside him under "DISASTER STRUCK", 78.

Neither beat the creator's own photo, which scored 80. And ChatGPT changed the dog: the creator's small black dog became a bigger, Labrador-like one, even though the prompt said to keep the dog exactly as it was. Two lessons: never ask for your subject small, and if your own photo already works, AI may not improve it.

ChatGPT or Gemini?

Across the 20 new pictures, neither won every time:

  • ChatGPT scored higher on the ciabatta brief (89 vs 81), the remake (90 vs 85), the face prompt (87 vs 78), and the Minecraft and bothy one-liners (84 vs 77, 83 vs 72).
  • Nano Banana scored higher on the Minecraft brief (82 vs 73), both bothy briefs (75 vs 69, 78 vs 77), the versus split (85 vs 84) and the ciabatta one-liner (78 vs 77).
  • Nano Banana gets the shape right. Every picture came back 1376×768. ChatGPT's came back 1536×1024 every time, so YouTube adds black bars unless you crop it, and the critique took points off for them.
  • Nano Banana is faster. It took 9 to 17 seconds a picture. ChatGPT took 24 to 90 seconds.

So pick the one you already use, follow the rules above, and crop ChatGPT's pictures to 16:9 with a thumbnail resizer.

Skip the prompt

Every good result above needed the right pictures, a careful description and a few rules. ThumbnailPeak's AI thumbnail generator writes that brief for you: paste your video's link, and it reads the title, the description and frames from the video, and asks for a 16:9 picture with a few words that aren't in your title. It can make three different ideas at once and put your face in them from a face library. It needs a free account, which comes with 12 credits; one thumbnail uses 3 credits, a batch of three 6.

Whichever way you make it, check it before you upload: preview it in a real YouTube feed next to the videos it'll compete with, and read every word.

Key takeaways

  1. The one-liner invents people. A baker and a hiker who aren't in the videos, from titles alone.
  2. Attach real pictures, and the right ones. Frames that don't show your title's promise can lose the idea.
  3. Remaking your current thumbnail scored best: 85 to 90.
  4. Never ask for your subject small. That brief scored 69.
  5. Read every word. "WHY SO EXPENSIVE?" became "WHY SO", and still scored 90.
  6. ChatGPT makes 3:2, Nano Banana 16:9. Crop ChatGPT's before you upload.

Tested on 1 November 2026. ChatGPT: OpenAI's API, a chat model (gpt-6-luna) with the image generation tool set to chatgpt-image-latest. Nano Banana 2: Google Cloud's Vertex AI, gemini-3.1-flash-image, with no aspect-ratio setting. One try per prompt, first picture kept. Frames are YouTube's own stills from each video. Scores are from ThumbnailPeak's critique with each video's title, on the pictures as YouTube shows them. The before-and-after scores are from our Nano Banana test the same day. The face photo is an AI-made stand-in, not a real person. ThumbnailPeak is not affiliated with OpenAI or Google. ChatGPT is OpenAI's trademark, Gemini and Nano Banana are Google's, and Minecraft is Mojang's.

The tests use three videos published under the Creative Commons Attribution licence (CC BY 3.0): "$1 vs $1,000,000 Tornado Shelter in Minecraft" by Luka Andrews, "Homemade Ciabatta That Tastes Better Than the Store!" by Skip The Junk and "Overnight ALONE with my Dog in a TINY Mountain Bothy: Things Didn't Go To Plan!" by Walk With Wallace. Their current thumbnails are shown as published. The AI pictures are adaptations made from their titles, frames and thumbnails, shown unedited, and the creators did not make or endorse them.

Get a thumbnail built from your own video

Paste a video link, or describe one you haven't uploaded yet, and AI designs a thumbnail from your video, then scores it.

Generate a thumbnail

3 credits · 12 free credits with an account

Keep reading

AI

Can Nano Banana Make a YouTube Thumbnail? We Tested Four Gemini Models

Read time:12 min read
AI

Can ChatGPT Make a YouTube Thumbnail? We Tried It Three Ways

Read time:9 min read
AI

Are AI Thumbnails Allowed on YouTube? The Rules, the AI Label and Copyright

Read time:12 min read