AI thumbnail that matches my channel style: the reference method

To get an AI thumbnail that matches your channel, show the model your own thumbnails and describe your look the same way every time. On Sume, send two or three of your best thumbnails as input_references to POST /v1/images and paste one fixed style paragraph into the prompt, so each new thumbnail starts from the same brief.

YouTube's Made on YouTube post, read 2026-09-29, lists channel-matched thumbnail generation among its new Studio tools without describing how it works. This is a method you can run today with your own files.

Which of my thumbnails should I use as references?

Pick a few that share the traits you want repeated and leave out one-off experiments. openai/gpt-image-2.5 accepts up to 16 image references; bytedance-seed/seedream-4.5 is listed with a maximum of 10. More is not better: three consistent examples say more than ten different ones.

What goes in the style paragraph?

Write down what the references show, in plain words, and reuse it unchanged. A single paragraph keeps a set of thumbnails consistent when the topic prompt changes.

How do I send it?

Reference URLs must be public HTTPS. Keep the size the same on every request; a custom image_size of 3840 by 2160 passes the model's rules of edges in multiples of 16, a 3840 maximum edge, and at most 8,294,400 pixels.

curl -X POST https://api.sume.com/v1/images \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "openai/gpt-image-2.5",
    "prompt": "New thumbnail in the style of the references: subject on the right third, warm yellow background, empty title space, no text. Topic: cold brew mistakes",
    "input_references": [
      { "type": "image_url", "image_url": { "url": "https://example.com/thumb-1.jpg" } },
      { "type": "image_url", "image_url": { "url": "https://example.com/thumb-2.jpg" } }
    ],
    "image_size": { "width": 3840, "height": 2160 }
  }'

How do I keep a set consistent over time?

Save the two or three reference URLs and the style paragraph in one file and change only the topic line for each new video. If a result drifts, add one more reference that shows what you wanted rather than making the paragraph longer.

Ask for empty title space and add words yourself in an editor, so every thumbnail has the same lettering. Generated text can be misspelled, and a mismatched font breaks the channel look you are trying to keep.

Will it match exactly?

No. A reference guides the result; it does not copy it, and a new image can differ in faces, logos and fonts. Compare each result with your references side by side, and set the title text in an editor so the lettering is the same on every video. Sume covers the image file only; upload it in YouTube Studio.