Rebounder Tech Blog

Written by the people who actually run these systems in production.

A Prompt Template for Mass-Producing Thumbnails

Published About 2 min readBy the Rebounder engineering team — the people who operate these systems

This article may contain affiliate links. Its content is not affected by advertising.

In short

To mass-produce thumbnails with image generation, template three things — subject large, title band at the bottom, tone fixed — and swap only the subject and title per article.

Make a template

Not starting over each time is the trick to producing them at volume. The skeleton is three things.

① Subject large — the article’s subject at about 60% of the frame, background minimal. The more elements you cram in, the less legible it becomes.

② A title band at the bottom — bold title text, large, along the bottom of the frame. Thumbnails display small, so specify a size that stays readable when shrunk, not a restrained band.

③ Fix the tone — keeping the colour and the mood consistent makes the articles look coherent when lined up.

Template those three and swap only “the subject” and “the title text” per article. A few minutes per image.

The limit: Japanese text breaks

Honestly: image generation is bad at text, Japanese especially. The title’s characters sometimes come out wrong.

Two responses.

  • Accept some breakage as character (prioritise speed)
  • For text that really has to be exact, add it on top with another tool after generating

I take the former.

Do not make fake photographs

There is one line I hold. Even for a hardware review’s thumbnail, I do not make a fake image that looks like a genuine photograph.

It stays firmly one image of decoration and concept. What it is actually like to use goes in the body. The thinking is that an image is an entrance, not evidence.

What having a template really does

More than the three components, the decision not to think it through again each time is what works.

The temptation is to think it takes time because you are bad at design, but what was actually eating time was decisions. The size of the subject, the position of the text, the colour — searching for the optimum per article means agonising every time. Fix it once and the things to agonise over drop to two (subject and title).

And consistency emerges as a result of fixing it, so quality goes up too. A rare area where not thinking each time gives the better result.

Frequently asked questions

Q1Does Japanese title text come out correctly?

It breaks. Image generation is bad at text, Japanese especially, and characters come out wrong. Two responses: accept some breakage and take the speed, or add the text on top with another tool after generating.

Q2Can it stand in for a photo of the real device?

No. The line we hold is: never make a fake image that looks like a genuine photo of the real hardware. A thumbnail is decoration, one image of a concept — not evidence for the article.

Q3Do I need to rethink the prompt every time?

No. Fix the skeleton and swap only the subject and the title text per article. A few minutes each. Keeping the tone consistent also makes the articles look coherent lined up together.

What this article is based on

  • Markdown file

Every claim in this article comes from the records above. The repositories we operate are private so we cannot link to them, but which file, which lines, and at which commit we read them is recorded for every article. Nothing here is written from guesswork.