FLUX 3 Image 503: read the JSON status before you retry

A 503 from the FLUX 3 Image endpoint is not automatically a retry. BFL's generate docs say a 503 may be retryable only after you check the status in the JSON body, and they list other errors that a retry will never fix: 422 for unknown fields, 400 for an oversized image and 402 for insufficient credits. Read the body first, then decide. FLUX 3 Image is not in the Sume image catalog today; check GET /v1/images/models for the live list. The catalog lists FLUX.2 Pro and FLUX.2 Flex.

What BFL documents

The table lists what the BFL page says about errors and job statuses. It gives no numeric concurrency limit, so do not hard-code one.

Why a blind retry costs you

The last row is advice, not a BFL statement: BFL lists the moderation statuses as terminal outcomes, so a blind retry of the same prompt has little reason to succeed.

• Polling a Pending or Generating job is safe; submitting the same prompt again starts a second job.
• A 402 retried in a loop only repeats the failure.
• Result links expire within 1 hour, so copy images to your own storage early.

How Sume job errors differ

Sume shapes failures differently. A synchronous call to POST /v1/images waits up to 30 seconds. If the job fails terminally inside that wait, you get a 502 with an error envelope that carries code, message, retryable and next_action. If the wait budget runs out, or you send mode: "async", you get a 202 job envelope and read the images from GET /v1/jobs/{id}/result. Branch on the status code, not on the body shape. Per the Sume docs, failed or cancelled generations are not billed.

A retry loop that reads the flag

This loop retries only when Sume marks the failure retryable. It uses a FLUX.2 Pro id from the Sume catalog, because FLUX 3 Image is not listed.

import os, time, requests

H = {"Authorization": "Bearer " + os.environ["SUME_API_KEY"]}

def generate(prompt):
    for attempt in range(3):
        r = requests.post("https://api.sume.com/v1/images", headers=H, timeout=60,
                          json={"model": "black-forest-labs/flux.2-pro", "prompt": prompt})
        if r.status_code == 200:
            return r.json()["data"][0]["url"]
        if r.status_code == 202:
            return "slow job, poll " + r.json()["data"]["status_url"]
        err = r.json().get("error", {})
        print(attempt, err.get("code"), err.get("retryable"), err.get("next_action"))
        if not err.get("retryable"):
            raise RuntimeError(err.get("message"))
        time.sleep(2 ** attempt)
    raise RuntimeError("still failing after 3 tries")

print(generate("a ceramic mug on a wooden table, soft window light"))