Turn a drawing into an image with GPT Image 2.5
Wikipedia says GPT Image 2.5 added Sketch, which turns drawings into images. Sume exposes GPT Image 2.5 with image references; here is how to send a drawing.

To turn a drawing into a finished image with GPT Image 2.5 on Sume, upload the drawing to a public HTTPS URL, pass it in input_references with a prompt that says what the sketch should become, and call openai/gpt-image-2.5 or the Sunburst variant.
What the model release says
According to Wikipedia's GPT Image page, OpenAI released GPT Image 2.5 on September 8, 2026. The page lists a Sketch feature that turns drawings into images, 50 percent lower generation latency, and two variants: Flare for speed and Sunburst for detail.
| Item | Value |
|---|---|
| Release | Sep 8, 2026 |
| Sketch | Turns drawings into images |
| Latency | 50% lower |
| Variants | Flare (speed), Sunburst (detail) |
What Sume documents
Sume lists 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 set to auto, transparent or opaque. Quality is auto, low, medium, high, xhigh or max, and omitted quality defaults to high.
The Sume docs do not describe a separate Sketch mode, so treat your drawing as a reference image and describe the transformation in the prompt. Reference URLs must be public HTTPS; localhost and private-network URLs are rejected.
curl -X POST https://api.sume.com/v1/images \
-H "x-api-key: $SUME_API_KEY" \
-H "Content-Type: application/json" \
-H "Idempotency-Key: sketch-to-image-001" \
-d '{
"model": "openai/gpt-image-2.5",
"prompt": "Turn this pencil sketch into a clean product render, soft studio light, keep the proportions.",
"input_references": [
{"type": "image_url", "image_url": {"url": "https://example.com/sketch.png"}}
],
"aspect_ratio": "auto",
"mode": "async"
}'Tips for sketches
- Use high contrast: dark lines on a plain background read best.
- Say what to keep (layout, proportions) and what to change (materials, light).
- Use Sunburst when detail matters and Flare for quick iterations.
- Use
mask_urlto limit the change to part of the image. - Poll the job and read the Sume-hosted signed URL in
data[].url.
Billing for image generation is all-or-nothing: a completed generation is billed in full and a failed or cancelled one is not billed.
Sources
Related posts
More in Use cases
- Voicemail drop: pick immediate or delayed detection
ElevenLabs added voicemail detection for Twilio calls with immediate and delayed modes. Record the message ahead of time with Sume's tts_create and drop it.
- Walmart rich media is restricted for alcohol, tobacco and firearms
Walmart Marketplace restricts rich media for alcohol, tobacco and firearms and may unpublish it. A category gate to run before you spend on AI clips.
- Avatar script CTA: no click the green button, WCAG 1.3.3
An avatar that says click the green button on the right fails WCAG 1.3.3 for some viewers. How to write the spoken call to action so it still works.
- Autoplay avatar welcome video with sound: WCAG 1.4.2 rule
Can an AI avatar welcome video autoplay with sound? WCAG 1.4.2 allows it only with a pause or volume control once audio runs past 3 seconds.
Written by Sume