FLUX 3 Image rewrites your instruction: result.prompt vs Sume's rows

FLUX 3 Image expands your instruction into a detailed prompt before it generates, and the expanded prompt comes back in result.prompt, according to BFL's edit docs. FLUX 3 Image is not listed in the Sume image catalog (read 2026-10-09). The listed Sume rows return no rewritten prompt, so log your own prompt per call.

What does BFL return?

In BFL's edit flow you send an instruction and optional boxes. For many edits the model also finds the element you named and adds a box for it. Both the expansion and any boxes it added come back in result.prompt. To control exactly which elements change, BFL tells you to use bounding boxes yourself.

Why should I store result.prompt?

Because the expansion is what the model actually saw. If an edit changed more than you meant, the expanded prompt shows which elements it thought you named. Copy it into your own request, edit it, and send it back, with explicit boxes if you want a tight edit. BFL also notes that changing the resolution makes a new generation, so a 2k rerun of a 768sq draft can come out with a different composition.

What do Sume's listed rows return?

The Sume image response has no revised_prompt or expanded-prompt field, and token counts are always 0. The revised_prompt post covers that gap on the GPT rows. Keep the prompt you sent in metadata, which Sume stores on the job and does not send to the provider.

{
  "model": "black-forest-labs/flux.2-pro",
  "prompt": "Replace the red sofa with a grey one. Keep the rest unchanged.",
  "input_references": [
    {
      "type": "image_url",
      "image_url": {
        "url": "https://example.com/room.jpg"
      }
    }
  ],
  "aspect_ratio": "auto",
  "metadata": {
    "prompt_version": "v3",
    "sent_prompt": "Replace the red sofa with a grey one. Keep the rest unchanged."
  }
}

What does it cost?

FLUX.2 Pro is $0.03 provider list per image, or $0.0375 billed after the 1.25 factor, on Sume. See the image models docs for the live endpoint line.

For auditing, store four things per call: the prompt you sent, the model id, the settings and the saved output file. That is enough to explain most surprises later. On a provider that rewrites prompts, add the rewritten text too. On Sume's listed rows, there is none to add, which makes your own log the only record.