FLUX 2 Prompting Guide: 8 Prompts With Settings and Cost, and Why a 1024×1024 Image Costs 2x 1024×768
A 1024×1024 FLUX 2 image costs $0.0276. A 1024×768 one costs $0.0138. The extra 262,144 pixels double the bill, and a 992×992 square avoids it. Here is how FLUX 2 reads a prompt, 8 prompts you can paste with the exact settings and cost of each, and the sizes that quietly bill 2 or 3 megapixels.
How FLUX 2 reads a prompt
Black Forest Labs, the lab behind FLUX, publishes its own FLUX.2 prompting guide. It is written for FLUX.2 [pro] and [max]; its reference page adds notes that cover every FLUX.2 variant. Six rules from it decide most results.
1. Order beats length. “FLUX.2 pays more attention to what comes first.” The priority BFL gives: main subject, key action, critical style, essential context, secondary details. Its framework is Subject + Action + Style + Context.
2. 30 to 80 words is the sweet spot. BFL suggests 10 to 30 words for quick concepts, 30 to 80 for most projects and 80 or more only for complex scenes.
3. There are no negative prompts. Say what you want: “sharp focus throughout,” not “no blur.”
4. Text goes in quotes. Quote the exact words, say where they sit, describe the font, size and color, keep the string short and mention it early.
5. Hex codes work when they are tied to objects. “The car is #FF0000” works better than “use red #FF0000 in the image,” in BFL’s own words.
6. Camera names are style controls. “Shot on Fujifilm X-T5, 35mm f/1.4” gives a more authentic photo than “professional photo.” Film stocks (Kodak Portra 400) and eras (“2000s digicam style,” “80s vintage photo”) work the same way.
8 prompts with settings and cost
Every prompt below runs on the FLUX 2 model page or through the queue API with the settings shown. Costs come from the price on that page: $0.0138 per started megapixel of output, times the number of images. We wrote these prompts for this guide following BFL’s rules; we did not render them for this post, so run them on a seed and judge for yourself.
1. Product shot: subject first, camera named
Black Forest Labs’ framework is Subject + Action + Style + Context, and FLUX.2 “pays more attention to what comes first.” A named camera, lens and aperture beat “professional photo.”
Prompt: Matte white ceramic pour-over coffee dripper releasing a thin stream of coffee into a glass carafe, product photography, shot on Sony A7IV, 85mm lens, f/5.6, soft window light from the left, light oak counter, clean minimal kitchen background
Settings: {"image_size": "landscape_4_3"} (1024×768, 1 megapixel)Cost: $0.0138
2. Brand colors with hex codes, in a square that bills 1 megapixel
Put “color” or “hex” before each code and tie every code to one object. BFL warns that “use #FF0000 somewhere” gives inconsistent results.
Prompt: A minimalist running sneaker on a concrete plinth, upper in color #1B6B6F, sole in color #F5E6D3, laces in color #E8A847, flat seamless background in color #F2EFEA, soft overhead studio light, product photography
Settings: {"image_size": {"width": 992, "height": 992}} (984,064 pixels, 1 megapixel)Cost: $0.0138 (the square_hd preset, 1024×1024, costs $0.0276)
3. A poster with exact text
Quote the words, say where they go, describe the type and its color. BFL’s text tips: front-load the text and keep strings short.
Prompt: Retro event poster with the headline "NIGHT MARKET" in bold condensed sans-serif letters in color #FF5733 across the top, the line "Fridays 6-11 PM" in small white text at the bottom, illustration of paper lanterns over food stalls, cream background, 70s screen-print style with visible grain
Settings: {"image_size": "portrait_4_3"} (768×1024, 1 megapixel)Cost: $0.0138
4. A JSON prompt for code that fills in the blanks
FLUX.2 reads a JSON object sent as a string in the prompt field. BFL recommends it for automation and for changing one element while the rest stays put.
import json, os, time, requests
BASE = "https://getmegabrain.com/api/gateway/v1"
HEADERS = {"Authorization": f"Bearer {os.environ['MEGABRAIN_API_KEY']}"}
prompt = {
"scene": "Studio product photography on a polished concrete surface",
"subjects": [{
"description": "Round cast-iron teapot with a bamboo handle, steam rising from the spout",
"position": "center foreground",
"color_palette": ["matte charcoal"]
}],
"style": "Ultra-realistic commercial product photography",
"color_palette": ["#2B2B2B", "#C8B79E", "#F2EFEA"],
"lighting": "Soft three-point softbox lighting, diffused highlights",
"composition": "rule of thirds, generous negative space on the right",
"camera": {"angle": "slightly high angle", "lens-mm": 85, "f-number": "f/5.6"}
}
job = requests.post(f"{BASE}/queue/flux-2", headers=HEADERS, json={
"prompt": json.dumps(prompt), # the JSON travels as a string
"image_size": "landscape_4_3", # 1024x768: 1 started megapixel
"seed": 42, # same seed + same request = same image
}).json()
while requests.get(job["status_url"], headers=HEADERS).json()["status"] not in ("COMPLETED", "FAILED"):
time.sleep(2)
print(requests.get(job["response_url"], headers=HEADERS).json())
# Next run: change only prompt["subjects"][0]["color_palette"] and keep the seed.Cost: $0.0138
5. A 9:16 cover for Reels and TikTok
Same prompt rules; the money is in the size. 1080×1920 is 2,073,600 pixels: 3 started megapixels. 1056×1872 is 1,976,832: 2.
Prompt: Close-up of a glass of iced matcha latte on a marble cafe table, condensation on the glass, morning sunlight through a window, shot on Fujifilm X-T5, 35mm f/1.4, generous empty space in the upper third for a caption
Settings: {"image_size": {"width": 1056, "height": 1872}}Cost: $0.0276 (1080×1920 costs $0.0414; the portrait_16_9 preset, 576×1024, costs $0.0138)
6. A film look
BFL’s style table: “shot on Kodak Portra 400, natural grain, organic colors” for analog, “2000s digicam style” or “80s vintage photo” for era looks.
Prompt: Empty seaside diner at blue hour, a red neon sign reading "OPEN" above the door, rain on the front window, warm light inside, shot on Kodak Portra 400, natural grain, organic colors, 35mm
Settings: {"image_size": "landscape_16_9"} (1024×576, 1 megapixel)Cost: $0.0138
7. Describe what you want, not what you don’t
FLUX.2 has no negative prompts. BFL: instead of “no blur,” say “sharp focus throughout”; instead of “no people,” describe an “empty scene.” Each re-roll is a new charge.
Prompt: Deserted city street at dawn, empty sidewalks, closed metal shutters on every shop, wet asphalt reflecting the last streetlights, sharp focus throughout, deep depth of field, f/11
Settings: {"image_size": "landscape_4_3"}Cost: $0.0138
8. Drafts on FLUX 2 Flash, the keeper on FLUX 2
FLUX 2 Flash takes the same prompt and the same image_size presets for less than half the price. Iterate on Flash, then run the final prompt on FLUX 2.
Prompt: Any prompt above, unchanged.
Settings: flux-2/flash, {"image_size": "landscape_4_3"}Cost: $0.0058 a draft (1024×1024 on Flash: $0.0115)
The 5 settings worth touching
FLUX 2 takes 9 fields and returns a 400 error for any other. The defaults are sensible; these are the ones to know.
| Field | Default | Range | What it does |
|---|---|---|---|
image_size | landscape_4_3 (1024×768) | 6 presets or custom, sides 512 to 2048 px | Sets the size, and with it the bill: per started megapixel |
seed | random | any integer | Same seed and same request give the same image: change one word and compare |
guidance_scale | 2.5 | 0 to 20 | How closely the image follows the prompt |
num_inference_steps | 28 | 4 to 50 | More steps: more detail, slower |
enable_prompt_expansion | false | true / false | Expands a short prompt automatically; leave it off when you need your exact words |
The other 4 are prompt, num_images (1 to 4, multiplies the price), output_format (png by default, or jpeg or webp) and acceleration (none, regular or high; higher is faster, and details may get simpler).
The size table: what each image actually bills
“Per started megapixel” means the pixel count is rounded up to the next million. The FLUX 2 playground shows it: Square (1024×1024) costs $0.0276, every other preset costs $0.0138. The custom sizes below follow the same rule; they must stay between 512 and 2048 pixels a side, and we keep them to multiples of 16, which BFL’s own API requires.
| Size | Pixels | Started megapixels | FLUX 2 | FLUX 2 Flash |
|---|---|---|---|---|
| 512×512 (square) | 262,144 | 1 | $0.0138 | $0.0058 |
| 1024×576 or 576×1024 | 589,824 | 1 | $0.0138 | $0.0058 |
| 1024×768 or 768×1024 | 786,432 | 1 | $0.0138 | $0.0058 |
| 992×992 (custom) | 984,064 | 1 | $0.0138 | $0.0058 |
| 1024×1024 (square_hd) | 1,048,576 | 2 | $0.0276 | $0.0115 |
| 1056×1872 (custom 9:16) | 1,976,832 | 2 | $0.0276 | $0.0115 |
| 1080×1920 (custom 9:16) | 2,073,600 | 3 | $0.0414 | — |
| 2048×2048 (maximum) | 4,194,304 | 5 | $0.0690 | — |
The 512×512 preset costs the same as 1024×768: both are 1 started megapixel. If you need a small thumbnail, generate at 1024×768 and resize it yourself. BFL calls 2048×2048 “4MP”; billed per started million pixels, it is 5.
Mistakes that waste money
Using square_hd for feed posts. 1024×1024 is 4.9% over 1 million pixels, and that 4.9% doubles the price. 992×992 loses 32 pixels a side.
Asking for 1080×1920 exactly. It bills 3 megapixels. Generate 1056×1872 for 2, or 576×1024 for 1, and scale up in your editor.
Writing “no text, no people, no blur.” No negative prompts: every re-roll costs another image. Describe the scene you want.
Long sentences of on-image text. BFL: “long strings are harder to render accurately.” Put the headline in the image and the paragraph in your layout.
num_images: 4 while you are still fixing the prompt. 4 images cost 4x. Find the prompt with 1 image on Flash, then ask for variants.
Editing a full-size phone photo. FLUX 2 Edit also bills every started megapixel of each input image. Our week of Reels from 1 product photo shows a 12 MP photo billing 13 megapixels before the output is counted.
What this means if you build with AI media
Put the size in code, not in a dropdown. If your app lets users pick “square,” map it to 992×992. If it makes vertical covers, map 9:16 to 1056×1872 or 576×1024. The model sees almost the same canvas; your bill sees 1 or 2 megapixels instead of 2 or 3.
Template the prompt as JSON. A fixed schema with a seed lets you swap the product, the color or the background while everything else stays the same. That is how BFL builds its own mug example step by step, one field at a time.
Write prompts in English, even for a non-English audience. FLUX.2 understands many languages, and BFL notes native-language prompts can make local scenes more authentic, but English prompts tend to be the most precise. Keep the quoted on-image text in the reader’s language.
Let your coding agent write the wrapper. Paste this into Cursor, Claude Code or Codex: “Write a Python function generate(prompt, aspect) that calls the FLUX 2 queue endpoint at https://getmegabrain.com/api/gateway/v1/queue/flux-2 with my MEGABRAIN_API_KEY, maps square to 992×992, 4:3 to 1024×768, 16:9 to 1024×576 and 9:16 to 1056×1872, polls status_url, saves the image and logs the cost at $0.0138 per started megapixel.” To run Claude Code or Codex itself on MegaBrain, see the Claude Code setup guide or the Codex setup guide.
Try it on MegaBrain
FLUX 2 costs $0.0138 per megapixel and FLUX 2 Flash $0.0058, with a playground on each page and the same queue API described in the media API docs. One API key and one pay-as-you-go balance also run Nano Banana, Seedream, Kling and the text models.
Sources
Pages opened November 17, 2026. Prompting rules come from Black Forest Labs’ documentation; prices and parameters from MegaBrain’s model pages; pixel counts and started megapixels are our arithmetic on them.
- Black Forest Labs — Prompting Guide, FLUX.2 [pro] & [max] (prompt structure, word order, prompt length, typography, hex colors, JSON schema, aspect ratios and resolution limits)
- Black Forest Labs — Prompt Reference (camera, lighting and style cheat sheets; FLUX.2 notes for all variants, including English prompts and reference-image limits)
- Black Forest Labs — JSON Structured Prompting (JSON sent as a string in the prompt field)
- Black Forest Labs — HEX Color Prompting
- MegaBrain — FLUX 2 model page (price per started megapixel, parameters and defaults)
- MegaBrain — FLUX 2 Flash model page
Full disclosure: MegaBrain sells pay-as-you-go API access to image, video and text models, including FLUX 2 and FLUX 2 Flash. MegaBrain is not affiliated with Black Forest Labs.
MegaBrain for Mac · an AI coworker
Give MegaBrain the task, not the steps.
It works on your files, writes documents, spreadsheets and decks, connects to your team’s tools and asks before anything risky. Free to start.
Newsletter
Stay in the loop
Get our latest AI research and guides. No spam, unsubscribe anytime.