FLUX.2 Prompting Guide: Control Layout, Text, and References

Prompt FLUX.2 with a clear subject, action, style, and setting; specify exact text and each reference image's role. Black Forest Labs supports natural language and JSON scene descriptions. 4, 6, 7

What are the key facts about FLUX.2 prompting?

How should you describe scenes, text, and references?

BFL's prompting documentation supports detailed scene descriptions, structured prompts, typography instructions, and reference-image editing. These provide several ways to state a creative brief precisely. 4, 5, 6

Start by describing a visible result: the main object, its action or placement, the surrounding scene, and the light. The practical aim is to give a reviewer requirements that can be checked in the output.

For example: “A blue ceramic teapot on the right side of a pale wooden table, photographed at eye level. Soft window light comes from the left. The wall on the left is plain and empty.”

This is an illustrative prompt, not a claim that a specific result has been generated. It defines the object, framing, and empty area separately, making it easier to identify which instruction needs revision.

For lettering, BFL advises specifying the intended words and their placement. Give a short line of text a clear role, such as a headline at the top of a poster or a label on a product. 5

For multiple references, identify what each contributes. BFL's guide recommends describing the role of each image, such as subject, style, or background, instead of leaving the model to infer the intended combination. 6

What can FLUX.2 prompts control?

FLUX.2's prompting guide says it does not support negative prompts. Describe the wanted scene positively: an empty tabletop, plain wall, or clearly visible face is a more direct instruction than a list of exclusions. 4

JSON formatting is an organizational tool for the prompt, not a separate guarantee of adherence. BFL documents structured prompting, but the image still needs to be reviewed against the fields you supplied. 7

Distinguish the prompt's content from the API request schema. A JSON scene description belongs inside the prompt string; output dimensions and reference-image fields belong in their documented request locations. 7, 8

For an infographic, supply checked wording and values yourself, then verify their rendering. Treat factual checking as a separate editorial task when reviewing the output.

For a layout with required lettering, inspect both spelling and placement. A correct headline in the wrong position can be just as unsuitable as a misspelled one for the intended design.

Reference preservation also needs review. BFL describes character consistency through image references, but your acceptance decision should still check the face, clothing, proportions, and other defining details. 9

When comparing prompts, hold the model and request settings steady. A change in endpoint, dimensions, or references makes it harder to decide whether a wording revision produced the useful difference.

How do you send a structured FLUX.2 prompt through the API?

1. Choose a named model and access route. The pro API is a useful documented starting point for the example below; use BFL's setup guide to create credentials. 2, 10

2. Write a short creative brief before formatting the prompt. Separate the subject, composition, light, required text, and any reference-image roles in your working notes.

3. Turn that brief into either natural language or a structured description. BFL documents both approaches; use the form that makes your requirements easiest to edit and review. 4, 7

4. Submit the prompt as a string. This example serializes a small scene description into the API's prompt field, while leaving output dimensions as ordinary API parameters. 7, 8
   

import json
import os
import requests

scene = {
    "subject": "A blue ceramic teapot on a pale wooden table",
    "composition": "Teapot on the right; open space on the left",
    "lighting": "Soft daylight from a window on the left",
    "text": "A small card beside the teapot reads TEA TIME",
}
response = requests.post(
    "https://api.bfl.ai/v1/flux-2-pro",
    headers={"x-key": os.environ["BFL_API_KEY"]},
    json={"prompt": json.dumps(scene), "width": 1024, "height": 1024},
    timeout=60,
)
response.raise_for_status()
print(response.json()["polling_url"])

The scene keys above organize this example's prompt; they are not additional required API fields. Read them as a written brief serialized into one string. 7, 8

1. Poll the returned URL and retrieve result.sample when the status is Ready. Save the request alongside the downloaded image so that each wording change has an identifiable result. 8

2. For editing, add the documented reference-image fields and name their roles in the prompt. For example, ask for the teapot from the first image in the room from the second, preserving its shape and glaze. 6, 11

3. Evaluate one requirement at a time. If the composition works but the card text fails, keep the composition description and revise the text instruction instead of rewriting the entire brief.

How do you revise a FLUX.2 prompt after a failed result?

Use a simple review note such as “object correct, framing correct, label wrong.” This keeps the next prompt focused and gives collaborators a shared reason for requesting another candidate.

For a poster, try deciding the text hierarchy before generating: headline, supporting line, and incidental lettering. Supply the exact words for each role and inspect them separately in the output.

For a reference edit, make a similar list of preserved properties. Decide whether shape, material, color, or identity is essential, and reject results that compromise the property that defines the task.

The FLUX photography guide offers broader scene-writing context. Use the sibling FLUX.2 family overview when choosing the model itself.

How do pro, flex, and Klein prompting controls compare?

Choose flex when you want to evaluate its extra controls for a specific design problem. Choose Klein when its deployment or workflow characteristics fit your project, then give it a fully specified prompt. 2

Keep your comparison grounded in the same brief. Score whether each required element appears correctly, and record the amount of revision needed, rather than judging only the most attractive sample from each model.

Should I use long keyword lists with FLUX.2?

BFL's FLUX.2 guide recommends a subject, action, style, and context. Add details that clarify the intended image rather than extending a keyword list without a specific purpose. 4

Can I use negative prompts?

FLUX.2 does not support negative prompts, according to BFL's prompting guide. Describe the desired visible scene directly, such as a plain wall or empty path. 4

Does a JSON prompt require a special endpoint?

A structured FLUX.2 scene description goes into the ordinary API prompt string. Keep controls such as width and height in their separate request fields. 7, 8

How do I combine references without confusing their roles?

For FLUX.2 reference editing, state what each image contributes: subject, setting, or style. BFL's multi-reference guide recommends assigning those roles explicitly. 6

Sources

• FLUX.2 launch and capabilities
• BFL model overview and controls
• Official FLUX.2 dev model card
• FLUX.2 prompting guide
• Typography and design guidance
• Multi-reference prompting guidance
• Structured JSON prompting documentation
• FLUX.2 text-to-image API
• Character and style consistency guidance
• BFL API-key setup
• FLUX.2 editing request format

Related guides on PromptZone

• Realistic Photos with FLUX
• Best SDXL Models in 2026