Thumbnails with AI

Why the thumbnail is critical

You can put ten hours into a video, and if the thumbnail is weak nobody clicks. Thumbnail and title together decide the click-through rate, and that rate feeds back into how much the platform shows the video at all. It is the highest-leverage hour in the whole production.

Click-through is a proxy, and chasing it alone backfires

This is the thing to understand before any advice about bright colours and shocked faces.

Recommendation systems are not trying to maximise your clicks. They are trying to keep people watching, and a click is only useful to them as a step towards that. So the signals get combined: a video that gets clicked and then abandoned tells the system that the click was misleading, and the distribution stops.

That produces a specific failure mode, and it is the one most creators walk into. A more sensational thumbnail lifts click-through, the video does not deliver what the picture implied, people leave in the first thirty seconds, and total views end up lower than with the honest thumbnail — while the dashboard shows an improved click rate the whole time.

So the real target is clicks that stay. The practical rule that follows: the thumbnail may promise anything the video actually delivers, and nothing else. That is not a moral position, it is what the system rewards.

Thumbnail and title are one unit

The single most actionable idea here, and the one most often missed: they are read together, in about a second, so they should say different things.

A thumbnail that repeats the title wastes half the space you were given. If the title is "I automated my invoicing and it broke three times", the thumbnail does not need those words — it needs the picture that makes the sentence interesting. Two words on screen and an image that raises the question the title answers, or the reverse.

A useful way to design the pair: decide which carries the subject and which carries the tension. Title says what it is, thumbnail says why you would care — or title poses the question and thumbnail shows the surprising result. Either split works; saying the same thing twice does not.

Write both at the same time, before you design. Most weak thumbnails are weak because they were made last, to accompany a title that had already used up the idea.

What has to survive one second at thumb size

Your thumbnail competes in a grid, on a phone, against a dozen others, at a size where fine detail simply does not exist. Three things can survive that, and you get to use about two:

• One clear subject, large, with a lot of separation from the background.
• Two or three words, very large, high contrast.
• One emotional or informational hook — an expression, a before-and-after, an object that does not belong.

Everything else is decoration that costs clarity. The most common improvement to an existing thumbnail is deletion: remove the logo, the border, the second subject, the drop shadow, the third colour.

Principles that hold up

• Contrast and clarity. One large central element that reads instantly at small size.
• Emotion in the face. If there is a face, a real expression beats a neutral one — though see the caution below about which face.
• Very little text. Three or four huge words at most, complementing the title rather than repeating it.
• Bold, limited colour. Saturated and warm stands out; busy backgrounds do not. Two or three colours, consistently.
• Curiosity with delivery. A hint that makes someone wonder — about something the video answers.

Where AI genuinely helps

• Backgrounds and objects — a dramatic scene or a specific object via image prompts, which is the most reliable use.
• Cut-out and composition — removing a background from your own photo and placing it on a generated scene, in one click.
• Variations — five concepts in the time it used to take to make one.
• Concepts — asking for thumbnail ideas given the topic, the audience and what the video actually proves.

Two things never to let the generator do

The text. Image models produce misspelled and malformed words, and a typo on a thumbnail is seen by everyone who does not click. Generate the image, then set the type in a real editor where you control the font, the size and the position.

The face, usually. Generated faces in thumbnails have become a recognised marker of low-effort content, and audiences now read them as a signal to scroll. They also cannot be your face, which is what builds recognition in a feed. Use a photo of yourself, badly lit if necessary, over a generated background — that combination is both more effective and cheaper than trying to generate a convincing person.

If you are running a faceless channel, lean on objects, screens, before-and-after pairs and text instead. A generated stock human is the weakest option available.

Prompting for a thumbnail, not for a picture

A thumbnail has requirements a beautiful image does not, and asking for the image you want produces something you cannot use. Four things change that.

• Ask for empty space where the text goes. "Subject on the left, plain uncluttered background on the right third." A perfectly filled frame leaves you nowhere to put words, and shrinking the subject to make room ruins the composition.
• Ask for separation. "Strong rim light on the subject, dark simple background." What makes a thumbnail work at small size is the subject popping off the background, and that is a lighting instruction, not a subject one.
• Ask for the right aspect ratio — wide, 16:9 — rather than cropping a square later and discovering the composition was centred.
• Ask for less. Image models fill frames with detail because detail looks impressive full-size. "Minimal, one object, plain background, no clutter" is closer to what survives a grid.

Then keep the prompt. Same lighting description, same background treatment, same palette, video after video — that repeated block is where the consistency described below actually comes from.

What to put in the frame when there is no face

Faceless channels and technical topics need a subject, and "a picture of the topic" is the weakest answer. Things that work at thumb size:

• Before and after, split down the middle. Instantly legible, promises a transformation, and delivers if the video shows one.
• The object at the centre of the story — the device, the screen, the physical thing — lit hard against a plain background.
• A screenshot with one element circled or arrowed. Cheap, specific, and it signals that the video contains something concrete.
• A number. Large, alone. "3 steps", "40 minutes", "£0" — a number reads faster than any word and implies the video is structured.
• Two things side by side with a versus between them, when the video really is a comparison.
• A red circle or arrow — overused and still effective, because it directs the eye in a grid where nothing else does.

What does not work: a logo, an abstract render, a generic stock scene, or the first frame of the video. All four are what people reach for when they have not decided what the video is about.

Being recognisable beats being striking

A one-off spectacular thumbnail wins a click. A consistent visual identity wins a subscriber's attention every time you publish, which is worth far more over a year.

Consistency here means a few fixed decisions you do not revisit per video: the same two or three colours, the same typeface and weight, text always in the same position, the same treatment of your face or subject. Someone scrolling should recognise it as yours before reading a word.

This is also what makes AI useful rather than chaotic in this workflow — the constraints are fixed and the generator fills them, exactly as described in the design guide. Without them you get twenty attractive thumbnails that share nothing and build no recognition.

The workflow

1. Write the title and the thumbnail concept together, before the video is edited if possible. If you cannot describe the thumbnail, the video may not have a clear enough idea yet.
2. Generate or shoot the elements — background, subject, any object.
3. Assemble in an editor with your fixed colours, typeface and positions.
4. Run the checks below.
5. Make a second version that differs in one deliberate way, for testing.

The four checks before you upload

• Thumb test. Shrink it on screen to about the size of a fingernail. If the subject and the words are not instantly clear, simplify — do not enlarge the text and call it done.
• Grey test. View it in greyscale. If the subject stops separating from the background, your contrast was coming from colour alone, which fails for a lot of viewers and on bad screens.
• Corner check. The platform overlays the duration in the bottom right and may crop edges in some placements. Keep anything that matters out of the corners and away from the edge.
• Context check. Put it next to the actual competing thumbnails for your topic — search the term and look. A thumbnail that is brilliant in isolation and identical to the five around it has failed at the only job it had.

Testing, done so the result means something

Platforms increasingly let you run several thumbnails and keep the winner, which is genuinely useful and is easy to misread. Generating variations is cheap with AI — the wider pipeline covers where it fits.

• Change one thing. Two thumbnails differing in every respect tell you which image won, not why, and nothing transfers to the next video.
• Let it run. Early click rates swing wildly, and a day-one winner is frequently the loser by week two.
• Check retention alongside it. If the winning thumbnail brought worse retention, it did not win.
• Write down the lesson, not the winner. "Faces outperform objects for this audience" is reusable; "version B won" is not.
• Do not re-test old videos endlessly. The gains are on the next video, where you apply what you learned.

Setting the words so they survive a phone

The text is where most home-made thumbnails lose, and the fixes are mechanical rather than artistic.

• One weight, very heavy. Thin and medium weights disappear at small size. Pick a bold or black weight and use only that.
• Give it an outline or a solid block behind it. Text over a photograph will sit on a light area in one part and a dark area in another; a stroke or a filled shape makes it legible regardless of what moved underneath.
• Two colours maximum across the whole thumbnail's text, and the same two every time.
• Upper case for two or three words, sentence case for anything longer — but if it is longer than three words, cut it instead.
• Keep it on one line where possible. Two short lines are acceptable; three is a paragraph and nobody reads a paragraph in a grid.
• Do not put it over the face. The expression is doing work; covering it wastes the strongest element you have.

A quick way to know if the type is big enough: at thumb size, the text should occupy an obviously large share of the frame — noticeably larger than feels comfortable while designing at full size. Almost every first attempt is too small.

Build a reusable set, once

The reason thumbnails feel like a chore is that most people start from nothing every time. An afternoon of setup removes that permanently and is what makes a consistent style survive a busy week.

1. A template file at the right dimensions, with your text positions, colours and typeface already in place. New thumbnail means duplicating it.
2. Three or four background prompts that reliably produce your look, saved as text.
3. A folder of cut-outs of yourself — five or six expressions, backgrounds already removed. Doing this once saves the cut-out step forever.
4. Your two colours and one typeface, written down, so a collaborator or a future you does not drift.
5. A checklist of the four checks above, pinned where you will see it.

After that, a thumbnail is fifteen minutes of composition rather than an hour of decisions — and the decisions were the part that made it feel expensive.

Where curiosity ends and clickbait starts

The line is not about how dramatic the thumbnail is. It is about whether the video delivers the thing the picture implied.

A bold expression over a real result is fine. A shocked face over a video with no surprise in it is not, and the punishment is automatic and quick — people leave, the system stops showing it, and the people who felt misled are exactly the ones who will not click your next one. Reputation with the algorithm is repairable; reputation with a viewer who felt tricked is less so.

A simple self-check: would a viewer who clicked feel the thumbnail was fair once the video ends? If you have to think about it, redesign.

Your back catalogue is the cheapest win available

Thumbnails are treated as something you make once, at upload, and then never touch. That leaves a lot of value on the floor.

An old video that still gets impressions is being offered to people and turned down. Replacing its thumbnail costs twenty minutes and no new production at all, and unlike a new upload you already know the video is worth watching — the retention data says so. Start with the videos that have the most impressions and the worst click rate; that list is usually short and obvious.

Two cautions. Give any change several weeks before judging it, since older videos accumulate impressions slowly. And do not touch a video that is performing well — a thumbnail people already recognise is an asset, and there is no version of this worth risking it for.

When the thumbnail is not the problem

• Impressions are very low. A click rate on a few hundred impressions tells you almost nothing, and no redesign creates demand for a topic nobody is looking for.
• Retention falls off a cliff. Then the thumbnail is working and the video is not. Fixing the picture will make it worse.
• The topic is the issue. A brilliant thumbnail on a video nobody wants is a well-designed advert for the wrong product.
• You have made three videos. There is not enough data yet to optimise anything. Make more, keep the style consistent, and look again at twenty.

Common mistakes

• Too much text. A sentence disappears at thumb size. Three words, enormous.
• Too busy. Competing elements produce a confused eye and no click. Delete something.
• Repeating the title. Half your space, wasted on words already on screen.
• Generated text or generated faces. Misspellings and a signal that reads as spam.
• Over-promising. The click you win is refunded with interest at the thirty-second mark.
• No consistent style. Recognition compounds; individual brilliance does not.