FLUX.2 prompt guide: structure, length and no negatives

Write FLUX.2 prompts as subject, then action, then style, then context, put the most important element first, and aim for 30 to 80 words. Describe what you want instead of what you do not: FLUX.2 has no negative prompts. Those are Black Forest Labs' rules for its [pro] and [max] models. On Sume you send the finished prompt to black-forest-labs/flux.2-pro.

The prompt rules come from BFL's prompting guide; the Sume side from the Image API docs. All read 2026-09-29.

How should I structure a FLUX.2 prompt?

BFL's framework is Subject + Action + Style + Context: the main focus, what it is doing, the artistic approach or medium, and the setting, lighting or mood. It adds that word order matters, because FLUX.2 pays more attention to what comes first, so lead with the subject and the key action and leave secondary details for the end.

Does FLUX.2 take a negative prompt?

No. BFL's guide says FLUX.2 does not support negative prompts and gives the workaround: instead of "no blur", write "sharp focus throughout"; instead of "no people", describe an "empty scene". Sume's Image API request table has no negative-prompt field either, so the advice carries over unchanged.

Which FLUX.2 controls does Sume pass through?

BFL's quick reference lists a seed, and guidance and steps for [flex]. Sume's docs list seed as part of the schema that no model advertises, so sending one returns 400 unsupported_parameter. They do not list guidance or steps at all, so leave them out and put the control in the prompt.

What does a structured call look like on Sume?

One black-forest-labs/flux.2-pro image is estimated at $0.04 on Sume. The prompt below follows BFL's order: subject, action, style, context.

curl -X POST https://api.sume.com/v1/images \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "black-forest-labs/flux.2-pro",
    "prompt": "A ceramicist shaping a bowl on a wheel, hands wet with slip, shot on a 35mm film camera, warm morning light in a small studio, sharp focus throughout",
    "aspect_ratio": "3:2"
  }'

What else is in BFL's prompting guide?

The guide also covers hex color codes, JSON structured prompts for complex scenes, camera and lens references for photorealism, and multi-language prompting. Its best-practice line for color is to tie a code to an object: "The car is #FF0000" instead of "use red #FF0000 in the image". Try each technique on one image before you build a template around it.

Do these rules apply to the other Sume models?

BFL wrote the guide for FLUX.2 [pro] and [max]. This post makes no claim that the structure or the length bands transfer to other models on Sume. Test a few prompts per model, and read GET /v1/images/models for what each one accepts.