GPT Image 2.5 edits through Sume: mask_url and 16 references

For GPT Image 2.5 edits, Sume exposes three controls in one request: up to 16 input_references, an optional mask_url, and background set to auto, transparent or opaque. OpenAI's guide says a transparent background needs png or webp output, so set output_format to one of those when you ask for transparency. The limits below are from the two pages. OpenAI's page says up to four input images where Sume documents 16, so test your own reference count.

What each side documents

An edit request

Sume rejects localhost, private-network and non-HTTPS reference URLs before submission. If a model's input_references range is 0 to 0, it is text-to-image only. Also set aspect_ratio: "auto" on edits to match the reference. Omitting it is not the same as auto.

curl -X POST https://api.sume.com/v1/images \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: mug-edit-001" \
  -d '{
    "model": "openai/gpt-image-2.5",
    "prompt": "Replace the label with a plain white label",
    "input_references": [
      {"type": "image_url", "image_url": {"url": "https://example.com/mug.jpg"}}
    ],
    "mask_url": "https://example.com/mug-mask.png",
    "aspect_ratio": "auto",
    "background": "transparent",
    "output_format": "png",
    "mode": "async"
  }'

Cost and timing

Flare and Sunburst share the same token rates in the Sume docs: $30 per million output image tokens, $8 per million input image tokens and $5 per million input text tokens, before Sume pricing. The docs give 1024x1024 output estimates of $0.09366 at xhigh and $0.21072 at max, before input tokens. Check the per-request estimate with the catalog pricing for your size and quality rather than computing it yourself.

Edits with high quality can run long. See GPT Image 2.5 can take 2 minutes: submit async for the async pattern.

Limits to remember

• mask_url is documented for GPT Image 2.5 only. Other models reject parameters their catalog descriptors do not list.
• Sume's transparent-still guidance for older paths points to Image 1.0 with transparency: true.
• The Sume docs do not describe how the mask is interpreted. Test with a known mask before you build a pipeline on it.

Planning the inputs

Count your references before you submit. Sume caps GPT Image 2.5 at 16, so a catalog of 20 product angles needs a selection step on your side. Host every reference and the mask at a public HTTPS URL, because localhost, private-network and non-HTTPS addresses are rejected before submission.

For size, both pages use multiples of 16 and a 3:1 ratio limit. A 2,048 by 1,152 image passes: 2,048 / 16 = 128 and 1,152 / 16 = 72, and its pixel count of 2,359,296 sits inside Sume's 655,360 to 8,294,400 range. Sume's 3,840 pixel maximum edge applies to both sides. Check any custom size against these rules before you pay for a request that will be rejected.