Make a YouTube Thumbnail

How it works

Your reference image goes to an image-editing model together with an instruction describing a thumbnail: one clear focal subject, strong contrast, readable negative space for text added later, an attention-grabbing but truthful visual hierarchy, and no generated text.

The model composes a new image at 1280 by 720. It is generating rather than cropping, which is why a second run gives you a different thumbnail instead of the same one again, and why the result resembles your reference without being it.

1280 by 720 is a legal size for the generator — both numbers are multiples of eight — so the output lands on those dimensions exactly with no correction step needed.

The size problem

This is the whole discipline of thumbnail design and almost nobody applies it.

You will make this image on a screen where it is perhaps 1280 pixels wide. It will be seen in a feed at around 210 pixels wide, on a phone, often in a scrolling column with a dozen others.

At that scale roughly a sixth of the linear detail survives. A face remains a face. An expression remains readable. Everything else — background objects, fine texture, thin lines, small text, secondary subjects — becomes undifferentiated texture.

So the only real test is to shrink the finished thumbnail to about 210 pixels wide and look at it. If you cannot immediately tell what it is about, it does not work, no matter how it looks at full size. Designers who do this routinely make better thumbnails than designers who are better at design.

Why one subject

Feed thumbnails compete for a decision made in well under a second.

A composition with one clear subject makes that decision easy: the eye lands, the subject registers, the viewer either wants it or does not. A composition with three subjects of similar weight makes the eye search, and searching is exactly what a scrolling viewer will not do.

This is also why the instruction asks for strong contrast. A high-contrast image holds its shape when reduced; a low-contrast one turns to mush, because the differences that defined its shapes were small to begin with and shrinking averages them away.

Leaving room for text

The instruction reserves negative space rather than filling the frame.

Almost every thumbnail ends up with a headline on it, and headline text needs somewhere quiet to sit. A generated image that fills every corner with interest leaves you overlaying words on detail, where they become hard to read and the detail underneath is wasted.

The text itself has to come from an editor.

And you have to check that the image did not bring its own. The instruction asks for no lettering, but asking is not the same as guaranteeing: a review on 13 August 2026 came back with PIAOF WILL CILALIET set confidently across the right-hand side. The model cannot spell and cannot reliably be told not to write, so treat any lettering in the output as something to crop or paint out before you use it.

Near-words on a thumbnail are exactly the error that survives review, because they look right at the size you are checking and fail at the size everyone else sees. Look at the full frame at full size, every time.

Publication gate

This page ships once the workflow has been run against a single-subject frame, a multi-subject scene, and a low-contrast source, with every output judged at 210 pixels wide as well as at full size.