Coffee roaster iced-coffee ad: GPT Image 2.5 still, then a clip

A coffee roaster can make an iced-coffee ad in two steps: generate one hero still with openai/gpt-image-2.5 on POST /v1/images, then pass that still as the first frame of a short clip on POST /v1/videos. Making the still first lets you fix the bag, the glass and the light cheaply before any motion is paid for.

The order matters because a bad still is a bad video. A still costs less to redo than a clip.

Step 1: the still

Sume lists ChatGPT Image 2.5 as openai/gpt-image-2.5 (Flare) and openai/gpt-image-2.5-sunburst. Both support text-to-image, up to 16 image references, an optional mask_url, and background. Quality accepts auto, low, medium, high, xhigh and max, with high as the default when you omit it.

Attach your real bag photo as an input_references entry so the label survives, and ask for the glass, the ice and the light. Use aspect_ratio of 9:16 if the clip is for stories, or 4:5 for a feed still (Sume documents 4:5 as Instagram portrait, 1080 by 1350).

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": "Iced coffee in a tall glass beside our coffee bag on a sunlit cafe counter, condensation on the glass, keep the bag label exactly as in the reference",
    "input_references": [
      { "type": "image_url", "image_url": { "url": "https://example.com/our-bag.jpg" } }
    ],
    "aspect_ratio": "9:16",
    "quality": "high"
  }'

Check the label

Image models can rewrite small text. Zoom into the bag label, the roast name and any weight on the pack, and compare them with your real packaging. If any letter is wrong, edit with a mask or regenerate, and do not move on to video with a wrong label in frame one.

A response of 200 is the image; 202 is a job envelope for slow configurations, which you then poll like any job.

Step 2: motion from the still

Take the finished still's media.sume.com URL and send it as frame_images with frame_type: first_frame to POST /v1/videos. Pick a model whose supported_frame_images includes first_frame: seedance-2 does in Sume's docs. Describe small motion, such as ice settling and condensation sliding, and keep the camera still.

This month's roundups point to short controlled clips and synced audio as the practical sweet spot for generated video (AI Video Generation Trends, read 2026-10-07). A 5 to 8 second loop fits.

Cost control

Read each model's price lines in its catalog (GET /v1/images/models and GET /v1/videos/models) before a batch. Iterate on the still at medium quality, then raise it for the final.