Pinterest Performance+ Creative picks by asset: feed it variants

Pinterest's June 17, 2026 announcement says Performance+ Creative uses a new AI model that picks creative at the asset level rather than the ad level, and reports 7.5% more click volume than its previous models. An asset-level chooser only helps if you give it distinct assets, so make a spread of product images and clips with Sume and queue them in a bulk run of up to 100 Format runs.

Everything about Pinterest here comes from its Cannes 2026 newsroom post, read on 2026-10-02; the 7.5% figure is Pinterest's own claim and the page does not describe a test design. Sume's side comes from Bulk runs and the Image API.

What did Pinterest actually announce?

The page names Performance+ Creative as a new AI model for dynamic creative selection that operates at asset level rather than ad level, with new ad review tools and more detailed creative reporting breakouts. It also lists a Business Assistant in closed beta in the U.S., a Pinterest MCP in alpha with partner agencies, and Ask Pinterest as a limited-access experiment.

The page gives no specs for asset counts, formats or file sizes, so check Pinterest's ad documentation before you build to a size.

Treat Pinterest's number with care: a reported lift against an older model says the new model is better than the old one, not that more assets will lift your own account.

Why does asset-level selection reward variety?

If the platform chooses among assets, ten near-identical images give it nothing to learn from. Variety should be in what the viewer sees: the angle, the setting, whether a person is present, the first frame of a clip. Keep brand rules fixed, and vary only the things a shopper would notice.

Write the matrix before you generate: five scenes times two crops is ten assets, each with an ID in the filename so the breakout report can be mapped back to the idea.

A short naming scheme pays off here. Something like sku, scene, crop and version, joined in a fixed order, lets you sort the asset-level report by idea. If a breakout shows that kitchen scenes beat white backgrounds for one SKU, you want to know that from the filename, not from opening forty files.

How do you produce the spread on Sume?

Two routes exist. For still images, the Image API makes several images per call with n, within each model's own ceiling, and accepts reference images so the product stays consistent. For a catalog of 40 SKUs, queue a Format run per SKU: a bulk run takes up to 100 items and a concurrency window, polls at GET /v1/format-run-queues/{queue_id}, and reads each child on its own.

A bulk request is a server-side queue of ordinary Format runs, one sandbox and one agent turn each, so it costs what the same runs cost one by one. The docs note service-account keys cannot create Format runs or queues, and your key needs the formats:write scope.

Run a small queue first, two or three items with concurrency of 1 or 2, to check the output before the full 100. A queue is cheap to start and awkward to recall: the docs say there is no public cancel-queue endpoint, only a cancel for each child run.

curl -sS -X POST "https://api.sume.com/v1/formats/chase/product-promo/bulk-runs" \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: pin-spring-set-001" \
  -d '{
    "concurrency": 3,
    "items": [
      { "instruction": "sku-101 lifestyle, kitchen", "input": { "url": "https://example.com/101.jpg" } },
      { "instruction": "sku-101 close-up detail", "input": { "url": "https://example.com/101.jpg" } }
    ]
  }'

What can Sume not promise here?

It cannot say variety will lift your clicks. Pinterest's 7.5% is a platform comparison against its earlier models, not a result for your account. Sume does not upload to Pinterest Ads Manager, read reporting or enforce Pinterest's creative policy.

Check the queue for a bad item before you launch; our posts on one bad item returns 400 and on a completed queue with failed items cover how a queue reports partial failure. Then read each output yourself: a generated product image can change a label or a logo.