Draft with GPT Image 2.5 Flare, finish with Sunburst: a two-pass edit

OpenAI's image guide recommends gpt-image-2.5-sunburst for editing precision and gpt-image-2.5-flare for fast everyday generation. A two-pass workflow follows from that: draft with Flare until the composition is right, then send the winning draft to Sunburst as a reference with an edit instruction.

On Sume the two are openai/gpt-image-2.5 (Flare) and openai/gpt-image-2.5-sunburst, both on POST /v1/images.

What each model is for

From the OpenAI guide and Sume's Image API page, read 2026-10-03.

The two-pass script

Pass one generates a draft. Pass two sends the draft URL as input_references with an edit prompt and aspect_ratio: "auto", which the Sume docs recommend on edit calls so the output matches the reference. Both calls handle the documented 200 versus 202 split: a 200 carries the image, a 202 carries a job to poll.

import os, time, requests

BASE = "https://api.sume.com"
H = {"Authorization": "Bearer " + os.environ["SUME_API_KEY"]}

def images(body):
    r = requests.post(BASE + "/v1/images", headers=H, json=body, timeout=60)
    r.raise_for_status()
    if r.status_code == 200:
        return r.json()["data"][0]["url"]
    job = r.json()["data"]
    while True:
        s = requests.get(job["status_url"], headers=H, timeout=60).json()
        d = s.get("data", s)
        if d.get("terminal"):
            break
        time.sleep(d.get("next_poll_after_seconds") or 5)
    res = requests.get(job["result_url"], headers=H, timeout=60).json()
    print(res)  # inspect the job result shape for the image URL
    raise SystemExit("read the artifact URL from the job result above")

draft = images({"model": "openai/gpt-image-2.5", "quality": "low",
                "prompt": "a ceramic mug on a wooden desk, morning light"})
final = images({"model": "openai/gpt-image-2.5-sunburst", "quality": "high",
                "aspect_ratio": "auto",
                "prompt": "keep the mug and light, make the desk dark walnut",
                "input_references": [{"type": "image_url", "image_url": {"url": draft}}]})
print(final)

Why a low-quality first pass

The point of pass one is composition, not finish.

• Sume's docs say auto quality reserves max, so name a tier explicitly. low for drafts and high for the final keeps the draft cheap to repeat.
• Slow settings (4K, high quality, large n) are the ones most likely to return 202 instead of an image, so the script handles both.
• Reference URLs must be public HTTPS. A Sume-hosted result URL qualifies; a localhost path does not.
• Change one thing per edit prompt. If the second pass drifts, the edit instruction was doing too much, not the model.

Limits of this recipe

The recipe follows the roles in OpenAI's guide, but neither page says Sunburst always beats Flare at edits on your content. Run the same edit on both models for five of your own images before you commit the second pass to a pipeline.