Pinterest A/B testing: two Pin creatives from one reference image

For a clean Pinterest A/B test of creative, make two Pin images that differ in exactly one thing, such as the background color or the angle, both from the same reference image. Sume's image API accepts reference images through input_references on models that allow them, so both versions can start from the same product photo.

Pinterest's announcement is read 2026-10-01; Sume details are from the Image API docs.

What did Pinterest add?

The newsroom post introduces self-serve A/B testing to make it easier to test campaign strategies and validate impact, across creative, targeting and within Performance+. It does not describe test mechanics, sample sizes or duration, so take those from Pinterest's own documentation.

What should differ between the two images?

One variable. If both the background and the headline area change, you cannot tell which one moved the result. Pick one row from the plan below and keep the rest of the prompt text identical.

How do I make both from one reference?

Send the reference in input_references and change only the line that differs. On edit calls the docs prefer aspect_ratio: "auto" to match the reference, and note that omitting the field is not the same as auto; or use image_size for exact pixels, such as 1024x1536 for 2:3. Give each request its own Idempotency-Key, and reuse a key only to retry the same request.

curl -X POST https://api.sume.com/v1/images \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: pin-ab-bg-warm-001" \
  -d '{
    "model": "openai/gpt-image-2.5",
    "prompt": "Same product on a warm neutral background, soft light",
    "input_references": [
      { "type": "image_url", "image_url": { "url": "https://example.com/reference.jpg" } }
    ],
    "image_size": "1024x1536"
  }'

Why not generate with n=2 instead?

n makes several takes of one prompt, so the two images would differ in ways you did not choose. Two controlled requests give you a known difference. The catalog's per-model n ceiling is also lower than the 10 the request schema allows, so read it before relying on a count. For ongoing variation inside a campaign, see asset-level variations in Performance+.