AI YouTube thumbnail maker: thumbnails and titles from your own video
OpenShorts makes YouTube thumbnails from the video they are for. It reads the transcript and a spread of frames, proposes 10 titles each paired with a short thumbnail text, finds frames where your face is large and sharp, and generates 1, 2, 3 or 4 different thumbnail concepts at once, each 1280x720 and under YouTube's 2 MB limit.
It is not a free web toy: it lives in the OpenShorts app (the YouTube Studio tab). On the hosted Cloud, thumbnail images are a paid-plan feature from $12/month, 3 minutes of the plan per batch. Titles and descriptions work on the free plan too, and the self-hosted edition includes everything under MIT with your own Gemini key.
Image generation is on paid plans, from $12/month (3 minutes a batch). The free plan gets titles, and self-hosting is free under MIT with your own Gemini key.
What the thumbnail studio does
In the app the feature is the YouTube Studio tab, and it is built around one idea: a thumbnail and a title are one package, so they should be made together and from the actual video rather than from a blank prompt. It runs in five steps, and you can stop after any of them.
- Input. Upload the video file, or type a title if you only want the thumbnail.
- Titles. 10 titles written from the transcript and frames, the strongest two marked as top picks, each with a 1 to 4 word thumbnail text that adds to the title rather than repeating it.
- Generate. Thumbnails built on that title and text, with your face, an optional background and your own instructions.
- Description. A YouTube description with a short hook, calls to action, chapters with timestamps and 5 to 10 hashtags.
- Publish. Upload the video to YouTube with the chosen title, description and thumbnail, through your connected account.
Everything it produces is editable before it is used: the title field, the thumbnail text, the instructions for the image and the description are all plain text boxes, and nothing is published until you press the button.
How to make a YouTube thumbnail with AI, in three steps
1. Upload the video, or type the title
Open the YouTube Studio tab in the app and drop in the video file. It is transcribed while you wait, ten frames are sampled across it, and Gemini writes 25 candidate titles before a second pass keeps the best 10, each paired with a 1 to 4 word thumbnail text. No video yet? Type your own title instead and go straight to the thumbnails.
2. Pick a title and generate thumbnails
Choose a title (or refine the list in the chat), then pick a frame of your face from the video or upload a photo, add optional instructions and a background, and generate 1, 2, 3 or 4 thumbnails at once. Each one is a different concept, cropped to 1280x720 and saved as a JPEG under 2 MB.
3. Check it on a phone, then download or publish
The phone preview shows the thumbnail at the size most viewers will see it. Download the one you want, or carry on: the studio writes a description with chapters from the transcript and can publish the video to YouTube with that title, description and thumbnail.
What comes out
| Property | What the studio produces |
|---|---|
| Size | 1280x720 pixels, cover-cropped to 16:9 whatever the image model returns |
| File | JPEG, re-compressed in steps until it is under 2 MB |
| Per batch | 1, 2, 3 or 4 thumbnails, each a different concept |
| Text | Set in Anton with a black stroke (Crisp), or painted by the model (AI painted) |
| Person | A frame of your face from the video, or an uploaded photo |
| Labelling | IPTC "trainedAlgorithmicMedia" in XMP plus an EXIF Software tag |
Each batch is designed before it is drawn. A text call plans every concept separately: the hook text, the side of the frame it sits on, the palette and the scene. Then one image call per concept runs in parallel, so three thumbnails are three ideas you can choose between, not three near-copies of the same picture. If none of them is right, change the instructions ("red and black, dramatic lighting, a laptop on fire") and press Regenerate.
YouTube thumbnail size and rules
YouTube's own guidance for custom thumbnails, checked 29 October 2026: a resolution of 1280x720 with a minimum width of 640 pixels, a 16:9 aspect ratio, an image format such as JPG, GIF or PNG, and a file under 2 MB. Custom thumbnails also need a verified channel; until your account is verified, YouTube offers only frames it picks from the video.
The studio meets the technical part by construction: every output is 1280x720, 16:9 and a JPEG under 2 MB, so it uploads without a resize. What no tool can do for you is the part that actually decides clicks, which is whether the image still reads when it is shrunk. A thumbnail is designed at 1280 pixels wide and mostly seen at a fraction of that, in a phone feed, a search result or the suggested column beside another video. The image prompts the studio sends ask for a picture that stays readable at 168x94 pixels: one subject, strong contrast, no clutter, no small details. The phone preview next to the results shows each candidate at that kind of size before you choose.
Crisp text or AI painted text
Image models are good at scenes and unreliable at letters. Ask one for a thumbnail that says "I QUIT MY JOB" and you will often get "I QUIT MY JOP", or the right words in the wrong language's accents. So the default, Crisp, splits the job: the model paints the scene and leaves a clean band (about 45% of the width, or 40% of the height) with no lettering at all, and the text is set afterwards in the Anton typeface with a heavy black stroke and a soft shadow. The words are exactly the ones you approved, spelled right, with accents intact.
AI painted lets the image model draw the text itself. When it works the lettering can sit inside the scene (on a sign, in neon, in the sky) in a way overlaid text cannot. When it does not, you get a typo in the largest type on your channel. It is there for the cases where the integration is worth a check.
Your face, taken from the video
Thumbnails with a recognisable creator tend to look like a channel rather than stock art, and the creator is already in the video. After analysis the studio samples frames, scores each one by how large and how sharp the face is (MediaPipe face detection plus a Laplacian sharpness measure), and keeps up to 5 spread across the runtime, so you get different moments rather than five copies of one second. Click one and it becomes the person reference for the batch. An uploaded photo of you, or of a guest who agreed to it, always takes priority over the frame. You can also upload a background image for the scene.
Titles written for the thumbnail, not next to it
The title pass reads the transcript plus ten frames at 1024 pixels, never the whole video. It runs twice. The first call brainstorms 25 titles across seven styles: a specific result, a curiosity gap that the video pays off, a contrarian take, a how-to with a concrete promise, a number or list, the question the viewer is already asking, and a story or confession. The second call acts as a critic: it scores each candidate for specificity, curiosity, clarity and honesty, drops near-duplicates of the same angle, and returns the best 10.
The rules baked into that pass are the ones that hold on a phone. The payoff has to land inside the first 50 characters, because that is roughly where mobile layouts cut a title, and nothing runs past 65. The main subject goes in the first three words. Titles stay in the language of the transcript. And each one comes with a thumbnail text of 1 to 4 words that complements it: if the title says what happened, the image says how it felt or what it cost, so the two together carry more than either repeats.
If none of the 10 fits, the chat beside the list takes plain requests ("more specific", "less clickbait", "mention the price", "write them in Spanish") and returns 8 new titles with their thumbnail texts. Up to the last 20 turns of that conversation are sent back each time, so it remembers what you already rejected.
Thumbnail tips that move click-through rate
Click-through rate is what a thumbnail is for, and most of what moves it is editorial rather than technical. The studio is built around these, but they apply whatever tool you use.
- One subject, one idea. A face, an object or a result. Three competing elements at 168 pixels wide become a blur.
- Few words, and not the title. One to four words that add something: the stakes, the number, the emotion. Repeating the title wastes the space.
- Contrast over detail. A bright subject on a dark ground, or the reverse. Fine textures and thin fonts disappear at small sizes.
- A readable face. Large, sharp and showing an expression that matches the video. Which is why the studio scores frames by face size and sharpness first.
- Keep the bottom-right corner clear. YouTube draws the video length there, over whatever you put underneath.
- Promise what the video delivers. A thumbnail that over-promises can win the click and lose the viewer in the first seconds, which shows up in retention and in what YouTube recommends next.
- Look like your channel. A consistent colour or layout lets returning viewers recognise your videos in a busy feed.
- Test, then decide. Where YouTube Studio's own thumbnail testing is available on your channel, upload two or three of the concepts from one batch and let the views decide, rather than guessing.
Which plan includes it, and what it costs
The studio is the same code in both editions; what differs is who pays for the Google calls behind it.
OpenShorts Cloud. Thumbnail image generation is a paid-plan feature: a free account that presses Generate is asked to upgrade. Every action is metered from the same minute balance your clips use, and the meter is deliberately light:
| Action | Minutes on Cloud |
|---|---|
| Analyse a video (transcription) | 1 per 10 minutes of source, never less than 2 |
| Title analysis on top of it | 1 |
| Refine titles in the chat | 0 |
| Generate a batch of thumbnails | 3 per batch, paid plans only |
| Write the description | 0 |
So a 20-minute video costs 3 minutes to analyse and 3 per batch of thumbnails after that. The free plan can run the analysis and get titles and a description from its own minutes; the images need a paid plan. The plans, from the pricing table:
| Plan | Price | Minutes a month |
|---|---|---|
| Cloud Free | $0, no card | First video up to 60 min, then 20 min. Clips watermarked |
| Cloud Starter | $12/month | 100, no watermark |
| Cloud Creator | $29/month | 300, no watermark |
| Cloud Pro | $59/month | 750, no watermark |
| Self-hosted | $0, MIT | No meter; your hardware and your own API keys |
Fair-use limits apply on Cloud: sources up to 180 minutes long, 40 title refinements per session and 60 title or description requests per hour per account. A free account that has run out of minutes can keep refining a session it already paid for, but cannot start a new one.
Self-hosted. The studio ships in the open source edition, free under MIT, with no minutes and no plan check. It needs your own Google Gemini key: titles and descriptions are text calls, thumbnails are image calls, and Google bills those to your key at its own rates. Publishing needs your own Upload-Post key and a connected YouTube profile.
Descriptions with chapters, and publishing
From the analysed video, the Description step writes a ready-to-paste YouTube description: a two or three sentence hook, the usual calls to subscribe, like and comment, a chapter list built from the transcript's timestamps (starting at 0:00, at least three chapters, each at least ten seconds apart, which is the format YouTube recognises as chapters), and 5 to 10 hashtags at the end, all under YouTube's 5,000-character limit and in the video's language. You can edit it in place or regenerate it. Writing it does not use minutes.
Publish then uploads that same video to YouTube as a new public video, with your title, the description and the thumbnail you picked, through your connected YouTube account. It runs in the background and the studio shows when the upload is done. If you only wanted the image, download it from the Generate step and stop there.
What it does not do
- It does not edit a thumbnail by hand. There is no layer editor; you steer it with instructions and regenerate, or open the JPEG in an image editor.
- It does not replace the thumbnail of a video already on YouTube. Publishing creates a new upload. For an existing video, download the file and set it in YouTube Studio.
- It does not take a YouTube link in the app. The studio works from an uploaded file (or from a title you type). Download your own video from YouTube Studio first if it is already online.
- It does not make vertical Shorts covers. Every output is 16:9 for regular videos.
- It does not generate images on the free Cloud plan. Titles and descriptions, yes; thumbnails need a paid plan or the self-hosted edition.
- It does not predict your CTR. The critic scores titles for clarity and honesty, but only your audience decides what gets clicked.
- It does not clear the rights to a face. Use your own face, or a guest's with their permission.
More tools for the same video
A thumbnail is one part of packaging a video. The free YouTube tag and title generator writes tags, title options and a description from a short brief, with no sign-up and no video needed. And once the long video is out, the same footage can feed your Shorts: turn it into YouTube Shorts with the clip generator, or see everything OpenShorts does with a YouTube video.
Common questions
- Is this AI thumbnail maker free?
- Not on the hosted free plan. On OpenShorts Cloud, generating thumbnail images is reserved for paid plans, from $12/month for 100 minutes, and each batch uses 3 of those minutes. The free plan (your first video up to 60 minutes, then 20 minutes a month, watermarked clips, no card) can analyse a video and get titles, but not images. The self-hosted edition is free under MIT and includes the studio; you bring your own Google Gemini key and Google bills the image calls.
- What size are the thumbnails?
- Exactly 1280x720 pixels, 16:9, saved as JPEG. The file is re-compressed in steps until it is under 2 MB, which is the limit YouTube puts on a custom thumbnail.
- Can it put my face on the thumbnail?
- Yes. After a video is analysed the studio scans sampled frames for a large, sharp face and offers up to 5 of them, spread across the runtime. Pick one and it is used as the person reference. You can also upload a photo, which takes priority over the frame.
- Is the text on the thumbnail spelled correctly?
- With the default Crisp setting, yes: the image model leaves one side empty and the text is set afterwards in the Anton typeface with a black stroke, from the exact words you chose. With AI painted, the image model draws the text itself, which can look more integrated but can also misspell it.
- How many thumbnails do I get per batch?
- You choose 1, 2, 3 or 4 in the studio (3 by default). They are different concepts, not one prompt repeated: a text call designs each one (hook text, which side the text sits on, palette, scene) before the images are generated in parallel. A batch costs 3 minutes on Cloud whatever the count.
- Does it write YouTube titles too?
- Yes, that is the first step. It brainstorms 25 titles across seven styles, then a critic scores them and keeps 10, with the payoff inside the first 50 characters and the main keyword in the first three words. A chat lets you ask for changes ("shorter", "less clickbait", "in Spanish") and returns 8 new options. Refining titles does not use minutes.
- Can I use it without a video?
- Yes. Choose "Write your own", type the title (up to 70 characters) and go straight to thumbnail generation. Without a video there are no face frames to pick from and no transcript, so the description is yours to write and publishing from the studio is not available.
- Does it change the thumbnail of a video already on my channel?
- No. Publishing from the studio uploads the video you analysed as a new public YouTube video, with the title, description and thumbnail you chose, through your connected YouTube account. To change an existing video, download the JPEG and upload it in YouTube Studio.
- Does it make Shorts thumbnails?
- It makes 16:9 thumbnails for regular YouTube videos. Shorts are vertical and YouTube handles their cover differently, so for Shorts the clip generator is the relevant tool, not the thumbnail studio.
- What languages does it work in?
- Titles, thumbnail text and descriptions are written in the language of the video's transcript, never translated unless you ask. In the title chat you can ask for another language and the new titles and thumbnail texts come back in it.
- Are the thumbnails marked as AI-generated?
- Yes, in the file metadata. Every generated JPEG carries the IPTC "trainedAlgorithmicMedia" source type in XMP and an EXIF Software tag naming OpenShorts. Nothing is drawn on the image itself.
Sources
- YouTube Help: Add video thumbnails (1280x720, minimum width 640, 16:9, JPG/GIF/PNG, under 2 MB; custom thumbnails need a verified account), checked 29 October 2026.
- Studio behaviour and limits:
thumbnail.py, the/api/thumbnail/*routes inapp.py,cloud/config.pyanddashboard/src/components/ThumbnailStudio.jsx. Source: github.com/mutonby/openshorts. - Plan prices and minutes: the pricing table, checked 5 October 2026.