FLUX 21:9 ultrawide image: FLUX.2 on Sume vs FLUX 3 ratios

Yes, you can make a 21:9 ultrawide image with a FLUX model on Sume: FLUX.2 Pro and FLUX.2 Flex list 21:9 and 9:21 among 13 aspect ratios, and POST /v1/images takes it as aspect_ratio. FLUX 3 Image, which launched on Oct 1, lists fixed ratios "from 21:9 down to 9:21" on Replicate, so the ultrawide endpoints are the same, but FLUX 3 Image is not in Sume's catalog and FLUX.2 does not accept auto.

This post covers which ratios work, how to ask for a banner or a phone-wallpaper tall frame, and what Sume tells you when you ask for a ratio a model lacks.

Which ratios does FLUX 3 Image list?

Replicate's page for the model says aspect_ratio is auto by default or a fixed ratio from 21:9 down to 9:21 (read 2026-10-03). OpenRouter's page says its listing has 15 options including auto and does not name them. I could not read the exact list on either page, so I only quote the two endpoints, plus the auto default that follows the first reference image.

Black Forest Labs' own overview pages did not give a ratio table in the text I could read, so treat the 21:9 to 9:21 range as the listed spread, not a complete enumeration.

Which ratios do FLUX.2 Pro and Flex take on Sume?

Sume's catalog code gives FLUX.2 Pro and Flex the same list: 1:1, 16:9, 9:16, 4:3, 3:4, 4:5, 5:4, 3:2, 2:3, 21:9, 9:21, 1:2, and 2:1. That is 13 values, and auto is not among them. Sume's docs say that a request that sets a value a model does not list is rejected with a 400 rather than silently changed.

A 21:9 banner request

Pin aspect_ratio to 21:9 and let Sume return a hosted URL. The prompt matters more than usual at this shape: wide frames invite a subject dead centre with empty sides, so say where the subject sits and what fills the rest.

import os
import requests

r = requests.post(
    "https://api.sume.com/v1/images",
    headers={"Authorization": f"Bearer {os.environ['SUME_API_KEY']}"},
    json={
        "model": "black-forest-labs/flux.2-pro",
        "prompt": "wide hero banner, a runner on the far left on a coastal "
                  "road at dawn, open sky filling the right two thirds",
        "aspect_ratio": "21:9",
    },
    timeout=60,
)
body = r.json()
if r.status_code == 200:
    print(body["data"][0]["url"])
else:
    print(r.status_code, body)

How do you write a prompt for a very wide or very tall frame?

At 21:9, name the subject's position (far left, lower third) and say what occupies the rest: sky, road, copy space. At 9:21, a phone-wallpaper shape, describe the vertical order from top to bottom, such as sky, subject, foreground, because the model otherwise stacks everything in the middle. If you will add text over the image later, ask for "empty copy space on the right" in words; Sume's FLUX.2 rows take a plain-text prompt and no layout coordinates.

FLUX 3 Image is the model with coordinates: BFL's box format places each element at a target box on a 0-1000 grid, so a banner layout can be specified numerically. On Sume the substitute is the prompt, plus a reference image of the subject when identity matters. The FLUX.2 pose and layout post shows how a reference image carries layout when words do not.

What resolution do you get at 21:9?

Sume does not publish a pixel size for a FLUX.2 request at any ratio, and the FLUX.2 rows have no resolution tier, so the only honest answer is to measure the file that comes back. If you need a specific banner size such as 2560 by 1080, plan a resize or a crop after generation, and keep a sharper source by upscaling rather than stretching.

What if you need a wider banner than 21:9?

FLUX.2's list stops at 21:9. Sume's generic docs list ratios such as 4:1 and 8:1, but the docs add that a model only accepts the values its own descriptors list, and Nano Banana is the family that carries the extreme banner shapes; the Nano Banana 2 banner ratios post covers those. Generate at the widest ratio a model gives you and crop, or outpaint; the FLUX outpainting note explains the canvas limits.

For Nano Banana Pro, the 21:9 ultrawide post compares Google's list with Sume's.

What about edits that should keep the source shape?

On an edit, auto matches the reference, but FLUX.2 does not list it, so pass the reference's own ratio from your list instead. Compute it from the file's width and height and pick the nearest listed value. If a reference is 2560 by 1080, that is about 2.37:1, so 21:9 is the nearest choice, and the result will be close to the source shape without being identical to its pixel count. The auto-ratio note lists which Sume models reject auto.