Sketch to image with GPT Image 2.5 API: send a drawing as a reference

To turn a sketch into an image with the GPT Image 2.5 API on Sume, host the drawing at a public HTTPS URL, send it in input_references, and describe the finished picture in the prompt, including "keep the layout of the sketch". Use aspect_ratio: "auto" so the output takes the drawing's shape.

This is the API route, not ChatGPT's own Sketch feature; I could not read OpenAI's announcement page, so this post makes no claim about that feature. Request facts come from Sume's Image API docs and OpenAI's model page, read 2026-09-29. A general walkthrough is in sketch to image AI.

What does the request look like?

Photograph or scan the sketch flat and evenly lit. Name the materials, light and background you want.

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": "Turn this pencil sketch into a finished product render of a brushed aluminum desk lamp on a white background. Keep the shapes and proportions of the drawing.",
    "input_references": [
      { "type": "image_url", "image_url": { "url": "https://example.com/lamp-sketch.jpg" } }
    ],
    "aspect_ratio": "auto",
    "quality": "medium"
  }'

How closely will the result follow my lines?

OpenAI's guide says the model may occasionally struggle to keep visual consistency across generations, and Sume's docs make no fidelity promise for references. Treat the sketch as a strong hint, and check proportions against the drawing. Ask for several tries with n, then keep the closest.

Which model id should I use?

Either 2.5 id accepts the same request. OpenAI's Sunburst page calls it its most capable model for image generation and editing; try it and compare with Flare (openai/gpt-image-2.5).