Same host face on every YouTube thumbnail with GPT Image 2.5

To keep the same host face across a series of YouTube thumbnails, generate or choose one approved image of the host, send it as the reference every time, and repeat the appearance constraints while changing only the scene and action. This is the character method in OpenAI's prompting guides, and on Sume it is a POST /v1/images call with openai/gpt-image-2.5 and input_references. Expect some drift and check every thumbnail.

OpenAI's method for the same character

The OpenAI Cookbook guide describes continuing a character as: same character, new scene and action, appearance must remain unchanged. The image prompting page says to reuse the generated character image, describe a new scene, repeat the appearance constraints, and say not to redesign the character.

Run it on Sume

Host the anchor image at a public HTTPS URL. Sume's docs say reference URLs must be public HTTPS, and localhost, private-network and non-HTTPS URLs are rejected before submission.

Use only a face you have the right to use: your own, or a host who has agreed. For a series, keep the appearance block in a variable and change only the scene line.

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",
    "quality": "high",
    "image_size": "1280x720",
    "n": 2,
    "prompt": "Same host as the reference image, new scene: standing at a whiteboard, pointing at it, surprised expression. Do not change the face, hairstyle, skin tone or proportions. Headline text, exactly once: \"EPISODE 12\". No other text.",
    "input_references": [
      {"type": "image_url",
       "image_url": {"url": "https://example.com/host-anchor.png"}}
    ]
  }'

Size, count and cost

1280x720 meets Sume's custom-size rules: both edges are multiples of 16, the area is 921,600 pixels, which is inside the 655,360 to 8,294,400 range, and the ratio is under 3:1. Check the target platform's own size and file limits before upload; the YouTube thumbnail upload limit post covers one of them.

n samples several candidates per call. Each completed image is billed in full, so n: 2 is two images. Pick the one that looks most like the host and discard the rest.

Check for drift, then fix by hand

OpenAI's image guide says the model may occasionally struggle to maintain visual consistency for recurring characters. Put the new thumbnail next to the anchor at thumbnail size and compare hairline, face shape and skin tone. If it fails, retry with a stronger preserve line or fall back to the anchor photo as the reference instead of a previous generation, which stops errors from compounding.

The API returns only the image. A final pass in your own design tool, for fonts and brand elements, is normal practice for a series.