Disclosure: we make ThumbnailPeak, which has its own thumbnail generator, and we scored everything with ThumbnailPeak's own critique. Every image and score below is from 1 November 2026. The video is DIY PRO's, published under a Creative Commons licence.
Yes, Nano Banana can make a YouTube thumbnail. We gave Google's four current Nano Banana models the same three messages we gave ChatGPT in October, for the same go-kart video. Here's what happened:
- It gets the shape right. All 16 pictures came back wide, 1376×768 or close to it, which YouTube shows without black bars. ChatGPT's came back 3:2 every time.
- It's fast. The cheapest model, Nano Banana 2 Lite, took 4 to 10 seconds. The others took 8 to 29 seconds. ChatGPT took about a minute.
- Asked plainly, it scored 61 to 81 in ThumbnailPeak's critique. ChatGPT's three pictures from October scored 81 to 86 on the same day.
- It made the same mistakes, and some new ones. Every picture pasted words from the title onto itself. Seven of the 12 added a grinning man nobody asked for. Without frames, none drew the kart green. Nano Banana 2.1 read "Kick $$$" as "KICK ASS!" three times out of three.
- One better-written message fixed most of it. With a few rules and the frames attached, all four scored 84 to 91, and the cheapest model made the best thumbnail in the test.
What Nano Banana is
"Nano Banana is the name for Gemini's native image generation capabilities," in Google's own words. It isn't one model. On 1 November 2026 Google offered four current ones:
- Nano Banana 2 Lite, the fastest and cheapest.
- Nano Banana 2, the one Google's Gemini page calls free.
- Nano Banana 2.1, the newest, which the Gemini app gives to people on a Google AI plan.
- Nano Banana Pro, which paid users can switch to with "Redo with Pro".
Which one you get in the free Gemini app isn't simple. Google's Gemini page says Nano Banana 2 is free. Its help pages say the image model follows the chat model: Nano Banana 2 Lite when the chat is set to Flash-Lite, and Nano Banana 2 on Flash or Pro. They also say most free prompts run on Flash-Lite. So we tested all four.
How we tested
We picked the same video as our ChatGPT test: DIY PRO's "$20 Rust Bucket into Kick $$$ Go Kart W/ Harbor Freight 212". A rusty go-kart frame becomes a bright green two-seater with a new Predator 212 engine.
We didn't use the Gemini app. We sent each message to the four models through Google's own developer platform (Google Cloud), which is where the model names above come from. The app adds instructions of its own and a visible watermark, so your results there may differ.
We sent the same three messages as before, one try each, and kept the first image:
- Title only. "Make a YouTube thumbnail for my video "$20 Rust Bucket into Kick $$$ Go Kart W/ Harbor Freight 212"."
- Title and description. The same, with the video's description pasted below it.
- Title and frames. The same, with four frames from the video attached: the rusty frame, the frame painted green, the engine and the steering wheel.
Then we scored each picture in ThumbnailPeak's thumbnail analyzer with the video's title. We also re-scored ChatGPT's three pictures from October the same day, because our critique changed models on 24 October and now scores lower overall.
What Nano Banana got right
The shape. Every picture came back wide: 1376×768, which is within 1% of YouTube's 16:9, or 1408×768 from Nano Banana Pro twice, which is a little wider and gets two thin bars. Google's guide says that without other instructions the models make square pictures or copy the shape of an attached image, but "YouTube thumbnail" was enough for all four to pick a wide one. ChatGPT, asked the same way, made 1536×1024 every time.
The idea. All 12 went for rust-to-finished, the before-and-after that DIY PRO used on the real thumbnail. That's the strongest idea for this video, and ChatGPT found it too.
Speed and price. Nano Banana 2 Lite took 4 seconds for the title-only picture. Through Google's developer tools a picture costs about US$0.034 on Lite and on Nano Banana 2.1, US$0.067 on Nano Banana 2 and US$0.134 on Pro. In the Gemini app it's free within limits that Google doesn't publish as a number.
What it got wrong
It pasted the title onto the picture
All 12 put the title's own words on the thumbnail, several of them the whole title. Nano Banana 2's title-only try has "$20 RUST BUCKET into KICK $$$ GO KART! W/ HARBOR FREIGHT 212" across the top, then "BEFORE: $20 RUST BUCKET" and "AFTER: KICK $$$ GO KART" underneath.
That's why it scored 61. The critique gave it 1 out of 4 for legibility and said: "Multiple long text blocks shrink to tiny lettering at 168×94." YouTube shows your title right under the thumbnail, so words that repeat it fill space twice.
It made up people
Seven of the 12 added a person: a grinning man in a cap giving a thumbs-up, a man in a helmet sitting in the kart, a small inset photo of a man. Our messages never mentioned a person, and none of the frames we attached shows a face.
If the man looks like you, that's luck. If you show your face on your channel, a stranger's face on your thumbnail is a problem. And if he looks like someone real, it's a bigger one: see what YouTube says about AI likenesses.
It drew the wrong kart until it saw the real one
With only the title, the four models drew an orange kart, a black-and-orange kart and two red ones. With the description pasted in, they drew black and red karts, mostly two-seaters: the description says "2 seater" but never says green. Only with frames attached did three of the four draw it green. Nano Banana Pro didn't draw a kart at all with frames: it pasted the frames themselves under the whole title.
That's the same lesson as ChatGPT: without your footage, an image model invents. Our critique can't catch it, because it reads the picture and the title, not the video.
It changed the title's words
Nano Banana 2.1 wrote "KICK ASS!", "KICKASS!" and "KICK-ASS!" on its three pictures. DIY PRO wrote "Kick $$$". Maybe that's what the dollar signs stand for, but it's the creator's call, and YouTube's advertiser guidelines count the thumbnail when they judge swearing. Nano Banana Pro went the other way and wrote "KICK $$", dropping a dollar sign, the same slip ChatGPT made.
It drew logos
Three pictures turned "Harbor Freight 212" into a badge, one of them a road sign with "HARBOR FREIGHT 212 FULL BUILD!". A brand's name drawn as a logo is something to remove before you upload.
One better message fixed most of it
We then wrote one message that does what the critique kept asking for, attached the same four frames, and sent it to all four models and to ChatGPT:
Make a 16:9 YouTube thumbnail (1280×720) for my video "$20 Rust Bucket into Kick $$$ Go Kart W/ Harbor Freight 212". The attached frames are from the video: draw the go-kart exactly as it looks in them, a bright green home-built frame with a Predator 212 engine. Show the rusty old frame and the finished green kart 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.
Every picture improved:
- Nano Banana 2 Lite: 91, up from 63–80. It drew one kart, rusty at the front and green at the back, under "BUDGET BEAST". That's the highest score in the test.
- Nano Banana 2: 85, up from 61–70.
- Nano Banana 2.1: 86, up from 79–81.
- Nano Banana Pro: 84, up from 72–80.
- ChatGPT: 83. It still came back 3:2, even though the message asked for 16:9.
Every one drew the kart green, used a few words that aren't in the title (Nano Banana 2 used four, not three), and left out the made-up man. These are single tries, and a second try can come out a few points either way. But a 15-to-24-point jump for Nano Banana 2 isn't luck: it's the difference between pasting a title and making a thumbnail.
If you use Nano Banana, do this
- Attach frames from your video. Screenshots of the subject are what got the right kart. Three or four are enough.
- Say what it looks like. "A bright green home-built frame" fixed what the frames alone didn't for every model.
- Ban the title's words. "Don't use any words from the title", plus a word limit. This was the biggest single fix.
- Say "no people who aren't in the frames". Otherwise you may get a stranger.
- Keep the bottom-right corner empty. YouTube puts the video's length there.
- Read every word before you upload, and remove logos. "KICK ASS!" and "KICK $$" are easy to miss.
- Check it at feed size. Preview it in a YouTube feed next to the videos it'll compete with, and run it through a critique for what to fix next.
Our prompt guide for ChatGPT and Gemini has more messages like this one, each tested and scored.
When a prompt isn't the easy way
The better message works, but you have to know what to write, find your frames, and describe your subject. That's the part ThumbnailPeak's AI thumbnail maker does for you: paste your video's link, and it reads the title, the description and frames from the video, plans the idea and asks for a 16:9 picture. It keeps your text short, makes three different ideas if you want them, and can put your own face in them from a library of your photos.
It needs a free account, which comes with 12 credits; one thumbnail uses 3 credits, a batch of three 6. Nano Banana in the Gemini app is free within Google's limits. If you're happy writing a message like the one above, it's a good tool.
Key takeaways
- Nano Banana makes YouTube-shaped thumbnails, fast. All 16 were wide; the cheapest took seconds.
- Asked plainly, it scored 61 to 81, mostly because it pasted the whole title onto the picture.
- It invents. A grinning man in 7 of 12 pictures, and the wrong colour kart until it saw the frames.
- It may rewrite your words. "Kick $$$" became "KICK ASS!" three times.
- One better message lifted every model to 84–91. Attach frames, describe the subject, ban the title's words, no extra people.
Tested on 1 November 2026 through Google Cloud's Vertex AI (global endpoint): gemini-3.1-flash-lite-image (Nano Banana 2 Lite), gemini-3.1-flash-image (Nano Banana 2), gemini-nano-banana-2.1 (Nano Banana 2.1) and gemini-3-pro-image (Nano Banana Pro), with no aspect-ratio setting, one try per message, first image kept. ChatGPT's pictures were made on 21 October 2026 through OpenAI's API, a chat model (gpt-6-luna) with the image generation tool set to chatgpt-image-latest; its better-message picture was made the same way on 1 November. All scores are from ThumbnailPeak's critique with the video's title on 1 November 2026, on the pictures as YouTube shows them. Prices are Google's published per-image prices for a 1K picture, read the same day. ThumbnailPeak is not affiliated with Google or OpenAI. Gemini and Nano Banana are Google's trademarks, and ChatGPT is OpenAI's.
The test uses "$20 Rust Bucket into Kick $$$ Go Kart W/ Harbor Freight 212" by DIY PRO, published under the Creative Commons Attribution licence (CC BY 3.0). The frames are from that video, and the thumbnails were made from its title, its description and four of its frames. All are shown unedited, and DIY PRO did not make or endorse them.
