Before/after ad from one packshot: how the Sume Format builds it

sume-before-after is built in two stages. Per its catalog example, it first makes a first frame with ChatGPT Image 2 that holds identity, framing, lighting and background steady, then animates a clean reveal with Seedance 2.5 reference-to-video. The point of the two stages is that the before and the after share a scene, so only the thing you are selling changes. Call it at POST /v1/formats/sume/sume-before-after/runs.

What does the shipped example ask for?

How do I brief it for a non-beauty product?

Name the one thing that changes and the one thing that stays. For a home-organisation gift, the before is a cluttered shelf and the after is the same shelf with your product. Keep the camera locked, because the reveal is only believable if framing is the same in both halves.

curl -X POST https://api.sume.com/v1/formats/sume/sume-before-after/runs \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: before-after-shelf-001" \
  -d '{
    "instruction": "9:16 before/after: same locked shot of a cluttered shelf, then tidy with the attached organiser",
    "attachments": [{ "type": "input_image", "image_url": "https://example.com/organiser.jpg" }],
    "generation_spend_cap_usd": 20
  }'

What are the limits?

• A before/after about a person's body or skin is a claim about results. The clip is generated and is not evidence, so do not present it as a real outcome or a customer's result.
• Ad platforms have their own rules for before and after images, particularly for health and beauty; read them before you run the ad.
• Seedance 2.5 reference-to-video accepts 4 to 30 seconds at 480p, 720p and 1080p; the Format chooses its own length, so read the result for it.
• A run that cannot finish comes back failed with unattended_blocked.

How do I check the two halves match?

Pause the clip at the last frame of the before and the first frame of the after. The wall, the light and the framing should be the same, and only the subject should differ. If the background moved, the reveal reads as a cut between two shots, which is exactly what the matched first frame is supposed to prevent. Rerun with a tighter brief about what must stay locked.

Can I continue a run that came close?

Yes. Send previous_run_id with a short correction, for example to keep the background fixed, and the Format continues the earlier run as another turn instead of starting fresh. The earlier run must be terminal and resumable, otherwise the call is refused with a code such as previous_run_not_terminal or previous_run_not_resumable.

What should I compare it with?

The before-and-after video API post covers building the reveal yourself from model endpoints. This Format is the shortcut when the packshot and a one-line brief are all you have.